Oceanic thermal structure mediates dive sequences in a foraging seabird

Abstract Changes in marine ecosystems are easier to detect in upper‐level predators, like seabirds, which integrate trophic interactions throughout the food web. Here, we examined whether diving parameters and complexity in the temporal organization of diving behavior of little penguins (Eudyptula minor) are influenced by sea surface temperature (SST), water stratification, and wind speed—three oceanographic features influencing prey abundance and distribution in the water column. Using fractal time series analysis, we found that foraging complexity, expressed as the degree of long‐range correlations or memory in the dive series, was associated with SST and water stratification throughout the breeding season, but not with wind speed. Little penguins foraging in warmer/more‐stratified waters exhibited greater determinism (memory) in foraging sequences, likely as a response to prey aggregations near the thermocline. They also showed higher foraging efficiency, performed more dives and dove to shallower depths than those foraging in colder/less‐stratified waters. Reductions in the long‐term memory of dive sequences, or in other words increases in behavioral stochasticity, may suggest different strategies concerning the exploration–exploitation trade‐off under contrasting environmental conditions.


| INTRODUC TI ON
Understanding the responses and adaptations of organisms to environmental variability has received growing interest in ecology, especially in the context of ongoing global change (Doney et al., 2012). The last IPCC Synthesis Report (IPCC, 2014) noted that marine ecosystems are the most vulnerable to climate change.
the distribution and dispersion of organisms (García Molinos et al., 2015), and/or changes in the timing of ecosystem-level processes (Durant, Hjermann, Ottersen, & Stenseth, 2007). These changes are expected to affect trophic interactions and, by extension, the structure and functioning of ecosystems (Doney et al., 2012).
However, due to the complexity of trophic pathways within marine ecosystems, and to the inaccessibility of marine environments, it is extremely difficult to monitor these changes. One approach to addressing these difficulties involves the study of upper-level predators, which are more sensitive to changes in the marine food chain (Hindell, Bradshaw, Harcourt, & Guinet, 2003;Ramírez et al., 2016).
In that sense, exploring variation in emergent patterns across individuals experiencing different ecological conditions can enhance our understanding of animal-environment interactions (Reynolds, 2015). It has been proposed that complexity in behavior should correspond to complexity in the environment in terms of resource distributions, such that in heterogeneous environments, behavioral sequences are predicted to be more variable and complex, or in other words less deterministic, to enhance prey encounter rates (Alados, Escos, & Emlen, 1996;Humphries, Weimerskirch, Queiroz, Southall, & Sims, 2012;MacIntosh, 2014;MacIntosh, Alados, & Huffman, 2011;Shimada, Minesaki, & Hara, 1995;Viswanathan et al., 1999). For example, we previously showed that variation in the temporal organization of little penguin (Eudyptula minor) dive sequences varied across areas with different bathymetric profiles, with penguins foraging in deeper waters producing less deterministic or more stochastic sequences (Meyer et al., 2017). Thus, measuring temporal organization in behavior sequences across environmental gradients can potentially highlight the oceanographic conditions that represent challenging habitats.
In this light, the characterization of fractal properties in animal-derived time series has emerged as a useful tool to model temporal patterns of animal behavior as a complex process ranging from stochastic, uncorrelated behavior, to deterministic, long-range dependent (auto-correlated) behavior that persists across measurement scales (see review in MacIntosh, 2014). The balance between the stochastic and deterministic elements in behavioral sequences may be a critical point in the emergence and maintenance of behavioral flexibility (MacIntosh, 2015;Reynolds, Ropert-Coudert, Kato, Chiaradia, & MacIntosh, 2015). The degree of complexity in these behaviors is shaped by both internal states (e.g., stress, disease; Alados et al., 1996;Cottin et al., 2014;MacIntosh et al., 2011) and external conditions (e.g., environment;MacIntosh et al., 2011;Meyer et al., 2017). Specific "complexity signatures," as they have been termed (MacIntosh, 2014), are predicted to optimize the efficiency of biological encounters (Alados et al., 1996;MacIntosh et al., 2011;Shimada et al., 1995). For example, animals might exhibit greater stochasticity in more challenging environments as described above or rely on more ritualized patterns when conditions are amenable to them (MacIntosh et al., 2011;Meyer et al., 2017). Moreover, the competing demands of exploration versus exploitation of resources appear to be a key factor underlying the emergence of fractal patterns in behavior, providing a way of interpreting variation in the scaling properties observed .
Under these challenging conditions (i.e., low SST, mixed water, and strong winds), we hypothesized that little penguins should exhibit greater stochasticity in the temporal organization of their diving behavior because the more persistent behavioral patterns expected to emerge under less challenging or more homogeneous conditions would be disrupted. Following Reynolds et al. (2015), we hypothesize that foraging sequences may lose their correlation structure if alternations between diving and surface times take on less characteristic or "noisier" patterns, for example, if penguins begin to prioritize exploration, which is freer to vary, overexploitation. To test this hypothesis, we used detrended fluctuation analysis (Peng et al., 1994;Peng, Havlin, Stanley, & Goldberger, 1995) to examine the temporal organization of diving behavior of little penguins in relation to sea surface temperature, water stratification, and wind speeds. We also examined other, more traditional diving variables (i.e., mean dive depth, mean foraging effort per day, foraging efficiency, and mean number of dives per day) in relation to these environmental factors. The study is based on an 11-year dataset focusing on little penguins at Phillip Island Nature Parks, Australia, and additionally tests for associations between these dive parameters and other variables such as penguin sex (Pelletier, Chiaradia, Kato, & Ropert-Coudert, 2014) and breeding stage (i.e., incubation, guard, postguard; Chiaradia & Kerry, 1999;Saraux, Robinson-Laverick, Maho, Ropert-Coudert, & Chiaradia, 2011).

| Long-term monitoring of foraging behavior
The study was conducted on the little penguin breeding colony at Phillip Island, Australia (38°21′S, 145°09′E, Figure 1), which contains 28,000-32,000 breeding adults (Sutherland & Dann, 2012). We collected diving data from 353 foraging trips over 11 breeding seasons from 2001 to 2012 (except 2003) during the three stages of the breeding cycle, that is, incubation, guard, and postguard. Postguard data were not available for the 2006 season. We used three different types of data-loggers recording depth at 1 or 2 s intervals (see details in Table 1). Birds were captured inside their nest boxes.
Loggers were attached to their lower backs with marine Tesa ® tape (Beiersdorf AG, Hamburg, Germany; Wilson et al., 1997). Upon their return from a single trip at sea, birds were recaptured in the colony and loggers were retrieved. Each attachment and removal of the loggers were completed in less than 5 min. All animal research protocol across all years was carried out in accordance with the Phillip Island Nature Parks Animal Experimentation Ethics Committee approval and a research permit issued by the Department of Environment, Land, Water and Planning of the state of Victoria, Australia.

| Diving parameters
After recovery, raw data were downloaded from the loggers and analyzed using Igor Pro version 6.22A (Wavemetrics Inc.). We used a purpose-written script to adjust the depth to zero when birds were at the surface between two dives and extract diving parameters from the depth data, including maximum diving depth, dive and postdive duration, and the number of vertical undulations in the bottom phase of the dive for each dive deeper than 1 m, which are generally indicative of prey pursuit (Kato, Ropert-Coudert, Grémillet, & Cannell, 2006; Ropert-Coudert, Kato, Wilson, & Cannell, 2006). A bottom phase was defined as the first and last time during a dive that the depth-change rate became <0.25 m/s .
To allow comparison between the different stages, we calculated the mean foraging effort per day as the total diving duration (in minutes) divided by the trip duration (in days). Similarly, we calculated the mean number of dives per day as the total number of dives performed over a trip divided by the trip duration (in days). The prey encounters per unit time used as proxy for foraging efficiency was calculated as the total number of vertical undulations divided by the total diving duration over the trip (Sala, Wilson, & Quintana, 2012).
Previous studies have shown that dive sequences from foraging penguins are best characterized as persistent, long-range dependent fractional Gaussian noise (Cottin et al., 2014;Le Guen et al., 2018;MacIntosh, Pelletier, Chiaradia, Kato, & Ropert-Coudert, 2013;Meyer et al., 2017;Meyer, MacIntosh, Kato, Chiaradia, & Ropert-Coudert, 2015). We used two versions of the DFA algorithm in our study. The first (DFA) uses included a linear detrending method in which linear regressions are applied within each window to remove the trends (Peng et al., 1995). We used this algorithm on the first-order integrated sequences, assuming the original diving data represented an fGn process (MacIntosh et al., 2013). The second (DFA b ) uses a bridge detrending method in which the first and last points within each window are "bridged" to create a trendline (Cannon et al., 1997). We used this algorithm on the second-order integrated sequences, which reflect an fBm process (MacIntosh et al., 2013).
The outputs of these variants (DFA and DFA b ), insomuch as they reflect Hurst exponent estimators for fGn and fBm, respectively, are theoretically related to one another through the following relationship: α fGn ≈ α fBm − 1 (Seuront, 2010). Their agreement provides an internal test of validation that the approach is appropriate.
Both methods were run using the package "fractal" (Constantine & Percival, 2014) in R statistical software v3.4.2 (R Development Core Team, 2017). As scaling is generally lacking at the smallest and largest scales, and to avoid any mathematical biases from their TA B L E 1 Number of penguins (number of males/females) and logger characteristics for each breeding season inclusion, we used methods provided in (Seuront, 2010) 2015) and Meyer et al. (2017). As little penguins are visual predators that forage diurnally (Cannell & Cullen, 2008), we only used active diving data (usually from dusk to dawn) and excluded nightly nondiving data. For multiday trips, we calculated α DFA for each day and averaged the daily values over the whole trip.
Daily satellite-derived wind speed (m/s) had a spatial resolution of 0.25° from NOAA (http://www.ncdc.noaa.gov/data-acces s/marin eocea n-data/blend ed-globa l/blend ed-sea-winds; Zhang, Bates, & Reynolds, 2006). Thermal structure in the water column was obtained for the entire study period from a oceanographic simulation model of daily temperature aggregates at 22 depth levels between 0 and 77 m, with a spatial resolution of 0.25°. This dataset was obtained from the Mercator Ocean GLORYS2V4 (1993GLORYS2V4 ( -2015 global ocean reanalysis for the Global Ocean and Sea Ice Physics project available on the European Union Copernicus Marine Service Information website (http://marine.coper nicus.eu/servi ces-portf olio/acces s-toprodu cts/). These data were also used as a proxy of thermoclines, defined as layers of water in which temperature changes more rapidly with depth than the warm layers above and cold layers below (Fiedler, 2010  we considered that a thermocline was present in the water column when the highest dT/dD exceeded 0.07°C/m. Where thermoclines were detected, we measured the depth of the thermocline, corresponding to the highest dT/dD. Based on this depth, we calculated the intensity of the thermocline as the difference between the average water temperature above and below the thermocline (Kokubun et al., 2010). The strength of the thermocline was then used to distinguish between weak (≤1.5°C) and strong (>1.5°C) thermoclines.

| Statistics
All statistical analyses were conducted in R v3.4.2 (R Development Core Team, 2017). Preliminary analysis and previous studies (Pelletier, 2013) have shown that SST and water stratification can change concomitantly in the study area. We first tested the effect of SST and wind speed on foraging complexity and diving variables; SST was then substituted by our proxy to the thermal structure of the water column. Models including both SST and thermoclines as predictor variables were not considered because their close correspondence, that is, high SST usually corresponded to strong thermoclines, would lead to considerable variance inflation in the models.
Thermocline strength was classified as absent, weak, or strong in the models according to the description above. The distribution of each response variable was checked using the "fitdistrplus" package in R We constructed (a) linear mixed-effects models (LMMs, "lmerTest" package in R; Kuznetsova, Brockhoff, & Christensen, 2016) to investigate variation in α DFA and (b) generalized linear mixed-effects models (GLMMs, "MASS" package in R; Ripley et al., 2017) to investigate variations in mean dive depth, mean foraging effort per day, foraging efficiency and mean number of dives per day, as a function of SST, wind speed and thermocline presence and strength, as well as bird sex for each stage of the breeding season (detailed results are provided in Appendix S1). As different loggers were used and some individuals may have been sampled more than once across years, logger type and individual were added to the models as random factors. Numeric predictor variables were scaled and centered (z-transformed) to simplify interpretation of the parameter estimates in the statistical model (Zuur, Ieno, & Smith, 2007). Models were validated by visual examination of histograms of the residuals to ensure homogeneity and plots of residuals versus fitted values to ensure homoscedasticity. Using Cook's distance and visual checking of outliers, we found no evidence of overly influential data points in our models (Fox, 1991). To test for multicollinearity, variance inflation factors (VIF) were calculated for each predictor variable with the package "fmsb" and only variables with values lower than 2 were retained in the models (Zuur, Ieno, & Elphick, 2010). We included a quadratic term for SST if that improved model fit. However, the quadratic term for SST did not explain additional variance in the model so it was removed in favor of better interpretation of the linear effect. All correlation tests were done using the Pearson method. Student's t tests were used to compare average SST and wind speed between each stage. Descriptive results are presented as means and standard errors (SE) unless otherwise specified, and we set the alpha level for all statistical analyses at 0.05.

| Oceanography
As breeding season progress from spring to summer, the mean SST recorded during the foraging trips increased from 14.44°C

| Behavioral organization and environmental parameters
Little penguins exhibited higher values of α DFA with higher SST, irrespective of the breeding stage (Table 2; Figure 5a). Higher SST was also linked with shallower dives (Table 2 Figure 6e) in the presence of stronger thermoclines, irrespective of breeding stage. The presence of a stronger thermocline did not affect mean foraging effort per day, and weak thermoclines were not associated significantly with any aspects of diving behavior observed in this study. Stronger wind speeds were marginally associated with diving variables, as it was only associated with lower values of α DFA during the guard stage for one of the model used (Table 2). Finally, males dived deeper than females during the guard stage, but also showed lower foraging efficiency and lower mean numbers of dives per day than females during incubation (Table 2).

| D ISCUSS I ON
Sea surface temperature and water stratification influenced the diving behavior of little penguins, in particular the temporal organization of their diving behavior. When exposed to lower SST and lessstratified waters, little penguins exhibited greater stochasticity in their foraging sequences, but also deeper diving activity, lower foraging efficiency, and smaller numbers of dives, irrespective of breeding stage. These findings suggest that little penguin prey are more dispersed and/or less abundant in colder and less-stratified waters, leading penguins to spend more time exploring the environment rather than exploiting prey during foraging sequences .
On the one hand, variation in SST is known to influence the abundance and spatial distribution of small pelagic prey. For example, Rhinoceros auklets (Cerorhinca monocerata) breeding in the Sea of Japan are highly dependent on the seasonal availability of Japanese anchovy (Engraulis japonicas), which is positively related to increases in SST (Takahashi et al., 2001). Similarly, the optimal F I G U R E 3 Results of detrended fluctuation analysis (DFA) of diving sequences, showing scaling exponents for each stage F I G U R E 4 (a) The mean sea surface temperature (2001−2002/2012−2013) for each breeding stage; (b) average sea surface temperature in the absence of a thermocline (AB), the presence of a weak thermocline (WT), and the presence of a strong thermocline (ST) in the water column offshore temperature range for little penguin prey-capture success (Carroll et al., 2016) seems to mirror the thermal peak of availability of sardines (Sardinops sagax; Agenbag et al., 2003;Nevárez-Martıńez et al., 2001). In our study, the maximum SST was in the optimal offshore temperature range for capture success observed by Carroll et al. (2016), so that the positive association between foraging parameters and increasing SST (i.e., more deterministic sequences, shallower dives, higher foraging efficiency and higher numbers of dives) can reflect an improvement in foraging conditions for these penguins (e.g., higher prey availability), although the situation could become detrimental to foraging success if SST continues to increase and move away from the optimal temperature noted by Carroll et al. (2016).
On the other hand, strong water stratification, an indication of the presence of a thermocline in the water column, could enhance prey availability in the foraging area by acting as a physical barrier to prey dispersion, creating an aggregation of marine life (McInnes, Ryan, et al., 2017;Ropert-Coudert et al., 2009). Here, with strong thermoclines in the water column, penguins showed more determinism (less stochasticity) in their foraging sequences, as well as greater foraging efficiency, shallower dives, and higher numbers of dives, than they did in the presence of weak thermoclines or in their TA B L E 2 Summary of the effects on diving variables (α DFA , mean dive depth, foraging efficiency, foraging effort per day (min), and number of dives per day) of sea surface temperature (SST), wind speed (WS), thermoclines, and sex during the three periods of the breeding season using LMM and GLMM statistics  (Meyer et al., 2017). The potential for thermoclines to act as a physical barrier to prey, and thereby mimic the influence of bathymetric structure on both prey distribution and diving behavior, is also consistent with previous results concerning the role of the thermoclines/water stratification in prey encounters (Boyd et al., 2015;Kokubun et al., 2010;Pelletier et al., 2012;Ropert-Coudert et al., 2009;Waggitt et al., 2018).
Our results suggest that both higher SST and the presence of thermoclines lead to increased prey availability in the foraging area, which is reflected in the higher foraging performance of little penguins under such favorable environmental conditions. Moreover, such situations are likely to occur toward the second half of the breeding season, in particular during the chick-rearing (guard and postguard) stages. Both of these stages are critical periods for the growth and development of chicks, so changes in foraging efficiency at this stage may have dramatic consequences on reproductive output (Chiaradia & Nisbet, 2006). Future studies should investigate how the absence/presence of thermoclines in the foraging area over a season may influence the breeding success at the colony level.
The link between the temporal organization of the diving behavior and thermocline strength seems to reflect a trade-off pitting exploration of the environment against exploitation of prey as previously proposed . Under this exploration/ exploitation trade-off, the greater stochasticity in sequences of foraging behavior in colder and less-stratified waters resulted in an increase in target depths or variance from one dive to the next, which is likely to enhance information gathering about prey locations with increasingly heterogeneous prey fields. This pattern would lead to deeper dives on average and induce variability in both dive and postdive durations, reducing long-range dependence in dive sequences and thus leading to greater stochasticity in behavior.
In this context, when exploration of the environment is favored, foraging efficiency and numbers of dives decrease. Conversely, warmer and strongly-stratified waters favor exploitation due to increased accessibility and predictability of prey (Kokubun et al., 2010;Pelletier et al., 2012;Ropert-Coudert et al., 2009). Under favorable conditions, individuals may favor exploitation over exploration, leading to greater determinism in sequences of foraging behavior but also shallower dives, greater foraging efficiency and smaller overall numbers of dives performed. This more persistent and periodic behavior in periods of prey exploitation could reflect the physiological limits that constrain both dive duration and postdive duration, such as oxygen reserves (Wilson, 2003) and lactic acid build-up (Butler, 2006 predictability of prey fields. This is consistent with the other results described in this study in which penguins exhibited higher stochasticity in foraging sequences in response to challenging environmental conditions. However, we only detected such effect during the guard stage and it could be explained by the fact we did not record extreme wind speeds (>14 m/s), representative of storm conditions (Dehnhard et al., 2013;Saraux et al., 2016), during this study.
Using a comprehensive dataset over an 11-year period, we showed how the temporal organization of the diving behavior in a seabird might vary according to intra-and interannual variability of environmental conditions. More stochastic foraging sequences may reflect more challenging environmental conditions (e.g., colder and less-stratified waters), where prey patches are less predictable and when a switch toward more exploratory behavior is needed to buffer such changes. This approach might aid future studies aiming to understand how animal movement is affected by the distribution of prey and the physical environment. One component these studies could also take into account is how individual and social environment characteristics, especially for group-hunting species such as penguins (Berlincourt & Arnould, 2014;McInnes, McGeorge, Ginsberg, Pichegru, & Pistorius, 2017), shape the temporal organization of the diving behavior in marine predators, and their decisions to explore or exploit the environment.

ACK N OWLED G M ENTS
We thank the continuing support from the Phillip Island Nature Parks and its staff. We are grateful for all the students and fieldworkers who assisted on the data collection over the 11-year period: J. Yorke, P. Fallow, I. Zimmer, C. Saraux, L. Pelletier, F. Amélineau.
We thank also R. Reisinger for his advices and help to extract temperature profiles. This work was supported by grants from the Japan

CO N FLI C T O F I NTE R E S T
None declared. Data curation (lead); writing-review and editing (supporting). Cedric Sueur: Supervision (lead); writing-review and editing (supporting).

DATA AVA I L A B I L I T Y S TAT E M E N T
Diving data and metadata are archived and available on Dryad Yan Ropert-Coudert https://orcid.org/0000-0001-6494-5300