Satellite-Linked Archival Tracking

Satellite-linked pop-up archival transmitting (PAT) tags (Wildlife Computers, Redmond, WA, USA) were deployed on foraging sharks (n = 21) off north-west Scotland and south-west England between May and August in 2001 and 2002. A detailed description of tag deployment and tracking data preparation protocols can be found in Sims et al. (2006). Briefly, PAT tags were attached to basking sharks by first approaching them from behind in a small vessel. Using a modified speargun harpoon, tags were placed at the base of the first dorsal fin and held in position by a small stainless steel T-bar dart with a monofilament tether connected to the tag (Sims et al. 2003, 2006). Tagging was conducted under licences from the UK Home Office, English Nature and Scottish Natural Heritage. Shark locations during the period of tag attachment were derived using light-based geolocation (GLS), corrected for sea-surface temperature (SST), with a calculated error radius of 75·5 ± 54·5 km (Sims et al. 2006). In order to account for this spatial uncertainty, we resampled possible locations (n = 10 per GLS-derived location) from within the mean radius of error (Fig. 1). Resampled presence positions falling on land were discarded and replaced. We also resampled presence positions (n = 10) in the initial (vessel dGPS, error radius <5 m) and final (Argos pop-up location, error radius <1 km) locations per track, for equal weighting of all presence positions. All locations derived from this combined data set were treated as near-surface presence positions in further analyses.

Random Walk Simulations

The use of presence-only, serially autocorrelated tracking data to infer habitat preference has inherent complications (Aarts et al. 2008; Warton & Aarts 2013). In order to account for regions of habitat accessible to, but not actively utilized by, tracked sharks, we used a randomization procedure (cf. Heithaus et al. 2006; Sims et al. 2006) to generate correlated random walk simulations (n = 1000 per shark, total = 7000; adehabitatLT package for r; Calenge 2006). Simulated tracks were generated per shark such that the total number of locations equalled the original track length, and step lengths and turning angles were derived from distributions in each original track. Simulations were permitted to approach, but not cross, land, were time-matched to original tracks and were constrained within a region defined by the bounding box surrounding all locations obtained across all individuals (Fig. 1; 45° to 61° N, −15° to 6° W; hereafter ‘study area’). This study area includes the UK and Irish continental shelf region, and the shelf-break system (Fig. 2). Locations derived from this simulated data set were treated as pseudo-absences for statistical analysis.

Environmental Data

Composite front maps (7-day, rolling by 1 day; Miller 2009) were prepared for the study area using SST data obtained via the Advanced Very-High Resolution Radiometer (AVHRR) sensor and ocean colour data obtained via the Sea-Viewing Wide Field-of-View Sensor (SeaWiFS; Local Area Coverage, LAC), mapped to the study area at 1·1 km resolution using Mercator projection.

Seasonal front frequency maps quantifying the percentage time in which a front was detected in each pixel of the study area, as a ratio of positive detections to the number of cloud-free observations, were generated for each tracking year (Miller & Christodoulou 2014). As >95% of all tracking locations were obtained during the main UK basking shark sightings season (May–October), we used 7-day composite front maps from this period of each year (2001, 2002) to generate the front frequency data sets (thermal front detection threshold = 0·4 °C; chl-a min. front detection threshold = 0·06 mg m⁻³). We also generated seasonal front frequency maps for the preceding year, to assess the influence of the previous year's conditions on habitat selection (Fig. 2).

Contemporaneous front metrics (front distance fdist, front gradient density gdens, front persistence pfront) were generated from composite front maps and time-matched to shark tracks (7-day, rolling by 1 day). Front distance (fdist) quantifies the distance from any location in the study area to the closest simplified front, using a custom simplification algorithm (P. I. Miller, unpubl. data). Front gradient density (gdens) is the result of applying a Gaussian smoothing filter (σ = 5 pixels) to a map of the mean gradient magnitude values. It is designed to provide a local neighbourhood average of frontal gradient, avoiding the discrete nature of individual detected front contours. Front persistence (pfront) is the fraction of cloud-free observations of a pixel for which a front is detected. Again, a Gaussian filter (σ = 5 pixels) was applied, to provide a local neighbourhood average of frontal persistence.

Thresholds for front detection (Single-Image Edge Detection, SIED; Cayula & Cornillon 1992) are often chosen arbitrarily, yet the magnitude of cross-frontal temperature change is likely to influence associations between marine vertebrates and fronts (Etnoyer et al. 2006). We therefore systematically varied the SIED threshold used in the preparation of thermal composite front maps, from 0·2 °C (minimum detectable owing to SST scaling in original imagery) to 1·0 °C, generating a set of time-matched front metrics at each threshold. Values were obtained for each of these metrics, plus SST and chl-a with no front detection, for each location of the full data set (presence, resampled presence, pseudo-absence), and used as predictor variables in subsequent statistical modelling.

Statistical Analysis

We carried out a use–availability analysis over two spatiotemporal scales: (i) seasonal associations with zones of frequent frontal activity and (ii) near real-time associations with contemporaneous mesoscale thermal and chl-a fronts. We used logistic regression within a generalized linear mixed modelling framework (GLMM, lme4 package for r; Bates et al. 2014) to obtain estimates of the influence of each of the predictor variables on the probability of observing a presence (individual as random effect; binary presence/pseudo-absence response; binomial errors with logistic link function). Owing to serial autocorrelation in both tracking data and simulated tracks, which violates the assumption of independence essential to the use of GLMM, we used a nonparametric bootstrapping regime to iteratively resample both the presence and the pseudo-absence data sets for each model fit (Scales et al. 2015). A total of 1000 presence and 1000 pseudo-absence locations, weighted as per the proportion of the complete tracking data set contributed by each individual, were subsampled from each individual data set for each iteration. Resultant presence/pseudo-absence data sets were then used to fit models over 1000 iterations.

We repeated this procedure using (i) seasonal front frequency metrics (thermal, chl-a) for both the season in which the sharks were tracked, and the preceding year, and (ii) 7-day contemporaneous front metrics (thermal, chl-a; distance to closest front fdist, frontal gradient density gdens, frontal persistence pfront), together with time-matched SST and chl-a values. All 7-day contemporaneous front metrics and SST were standardized across the entire presence/pseudo-absence data set prior to the modelling procedure, by subtracting the mean and dividing by standard deviation (Zuur, Hilbe & Ieno 2013). This enables comparability of effect sizes between variables that are scaled differently in their original form. The distribution of chl-a was highly skewed, with a large predominance of small values. We therefore removed all spurious outlying values (>20 mg m⁻³) and transformed the resulting data set using a log₁₀ transformation to generate an explanatory variable with a distribution approaching normal.

Owing to colinearity between predictor variables, which was detected using pairwise plots and generalized variance inflation factors (GVIF; Zuur, Hilbe & Ieno 2013), each variable was fitted via maximum likelihood estimation as a standalone explanatory term in separate model runs (1000 iterations per term). Parameter distributions generated by each set of model iterations were used to obtain the mean and standard deviation of model intercepts, regression coefficients and standard errors of fitted terms, deviance explained and chi-square statistic and P-value from a likelihood ratio test against a null model with no fixed effects (with Restricted Maximum Likelihood; Table S1, Supporting information). Confidence intervals (CIs; 95%) were also calculated for each of the parameter distributions. Mean values and CIs of regression coefficients were plotted and used to assess the influence of each term on the probability of shark presence (CIs overlapping zero indicates non-significant term). To assess the influence of thermal gradient magnitude on the strength of associations with fronts, we repeated this modelling procedure for each set of time-matched metrics derived using different front detection thresholds (0·2, 0·4, 0·6, 0·8, 1·0 °C).

Satellite-Linked Archival Tracking

Of the 21 basking sharks tagged, sufficient data to reconstruct tracks were received from seven individuals (body length range 2·5–7·0 m), which were tracked for a cumulative total of 964 days, ranging from 72 to 213 days per individual. A total of 186 light-level geolocations were obtained (0·2 ± 0·05 per day) during this period. Associated dive data indicated that all sharks spent a significant proportion of this time foraging at the sea surface (Sims et al. 2006).

Seasonal Front Frequency

Basking shark tracking locations were clustered within broad-scale regions of high seasonal front frequency, in both SST and chl-a fields (Fig. 2). Logistic regression revealed that the probability of shark presence was higher in regions of frequent or persistent frontal activity (‘frontal zones’) during the basking shark surface sightings season (May–October) over 2 years (Fig. 3; Table S1). Thermal front frequency had a stronger influence over the probability of observing a presence than chl-a front frequency, although both contributed significant explanatory power to models (Fig. 3c; Table S1). The proportion of deviance explained was also found to be higher for thermal front frequency than for chl-a (thermal = 8·25 ± 2·32; chl-a = 1·65 ± 1·06).

Seasonal front frequency in the preceding year also had an influence on the probability of observing a presence (Fig. 3; Table S1). Model intercepts and regression coefficients were similar when modelling the influence of front frequency from the contemporaneous year and from the preceding year on shark presence (Table S1). Interannual variability in front frequency was low in both thermal and chl-a fields between 2000 and 2002 (Fig. 2, Table 1). We also observed a high degree of spatial correlation between the thermal and the chl-a seasonal front frequency metrics in each year (mean = 0·523 ± 0·04; 2000 = 0·476; 2001 = 0·561; 2002 = 0·533).

Time-Matched Front Metrics

Shark presence locations were significantly more likely to be associated with contemporaneous thermal and chl-a fronts than pseudo-absences derived from random walk simulations (Fig. 4; Tables S2 and S3, Supporting information). Distance to closest chl-a front (fdist) and all 7-day thermal front metrics (distance to closest simplified front, fdist; frontal gradient density, gdens; front persistence, pfront; 0·4 °C front detection threshold) were significant predictors of shark presence. Shark presence was more likely to be observed in closer proximity to thermal and chl-a fronts, at higher thermal gradient densities and in association with persistent thermal fronts than pseudo-absences. Indeed, some individuals appeared to spend days to weeks tracking the surface profile of strong thermal fronts, presumably foraging on aggregated prey (see Video S1, Supporting information).

Overall, 7-day chl-a front metrics held less explanatory power than thermal metrics, while distance to closest simplified chl-a front fdist explained a significant proportion of deviance, gdens and pfront had a less pronounced effect on the probability of shark presence (Fig. 5; Table S3). In addition, confidence intervals of the distribution of regression coefficients from bootstrapping approached zero for chl-a gdens and overlapped zero for chl-a pfront (Fig. 5). We can surmise that shark presence positions are more likely to be observed in closer proximity to chl-a fronts than pseudo-absences, but that chl-a gdens and pfront metrics have a lesser influence on probability of shark presence, presumably as a result of the ephemeral nature of chl-a blooms at fronts, and the spatial smoothing involved in preparation of these metrics. These results indicate that time-matched thermal front metrics are more useful predictors of shark presence than comparable chl-a metrics in this case.

Varying the thermal front detection threshold had a considerable effect on the magnitude of the logistic regression coefficient for the thermal fdist metric (Fig. 6; Table S2). Effect size and proportion of deviance explained increased with a higher detection threshold. Shark presences were more likely to be associated with stronger thermal fronts (1·0 °C cross-frontal temperature difference or ‘step’) than weaker features (0·2 °C difference), although all detection thresholds resulted in significant predictors (Fig. 6; Table S2). In contrast, altering the detection threshold had little influence over the effect sizes of the gdens and pfront metrics (Table S2), most likely as a result of the inclusion of the cross-frontal gradient in the gdens metric, and the tendency of fronts with a stronger cross-frontal gradient to persist through time (Bakun 2006).

Comparison with Standard SST and Chl-a Fields

Chlorophyll-a concentration was found to have a significant effect on the probability of shark presence, with log₁₀-transformed chl-a concentration explaining the highest proportion of deviance across model iterations (Fig. 5b; Table S3). Chl-a had a strongly positive effect as a predictor of shark presence, indicating that foraging habitat selection is tightly coupled with primary productivity. SST was also found to be a significant predictor, although this variable explained a considerably lower proportion of deviance than chl-a and time-matched front metrics, having a weak negative effect on the probability of shark presence (Fig. 5; Table S3).

Associations with Seasonally Persistent Frontal Zones

Seasonal front frequency, that is the number of times a front was detected in any one pixel (1·1 × 1·1 km) of the study area over the main UK basking shark surface sightings season (May–October), was found to be a significant predictor of shark presence for both thermal and chl-a frontal activity. Presence locations of tracked sharks were more likely to be found in association with seasonally persistent frontal zones than in other regions of the study area, although thermal front frequency was found to have a stronger effect than chl-a, perhaps owing to the propensity of thermal fronts to manifest in similar locations more frequently than chl-a fronts over the season (Kahru et al. 2012).

Furthermore, seasonal front frequency metrics from the preceding year were significant predictors of shark presence. Low interannual variability in the spatial extent of these persistent frontal zones over the study period (2000–2002) indicates that sharks may return to spatiotemporally predictable foraging grounds, in which they have previously experienced profitable prey encounter rates. Although we only have tracking data from seven different individuals tagged over two successive years, and none spanning 2 years, and so cannot determine whether the same sharks could be returning to forage in previously profitable regions, we can surmise that predictability of foraging hotspots is likely to be high over seasonal timescales. Basking sharks, like many pelagic marine vertebrates, may optimize foraging efficiency through orientation to the same broad-scale regions to search for suitable foraging areas, then using search patterns consistent with optimal random searches (Sims et al. 2008; Humphries et al. 2010) and more proximate clues to locate prey aggregations nested within (Cotton et al. 2005; Sims et al. 2006; Siders et al. 2013). Many marine vertebrates exhibit broad-scale foraging site fidelity over seasonal, annual or interannual timescales (e.g. seals, Bradshaw et al. 2004; sharks, Pade et al. 2009; Queiroz et al. 2012; whales, Irvine et al. 2014; seabirds, Patrick et al. 2014), indicating that spatio-temporal predictability of prey encounter rates influences habitat selection across taxa (e.g. seabirds, marine mammals; Weimerskirch 2007; Bost et al. 2009).

Spatial correlation between the locations of thermal and chl-a frontal zones with which sharks associate was also found to be high within the study area, over the 3 years’ of remotely sensed data analysed for this study. The locations of thermal and chl-a fronts often coincide (Le Fevre 1986; Belkin, Cornillon & Sherman 2009), since chl-a fronts frequently manifest where convergent processes occurring around thermal discontinuities aggregate nutrients and plankton in productive regions with high-background chl-a concentrations, such as at the peripheries of plankton blooms (Le Fevre 1986; Kahru et al. 2012). Although these mechanisms are not yet well understood, objective detection of regions of frequent frontal activity in both thermal and chl-a fields, such as that presented here, could aid in the identification of biophysical hotspots. Persistent thermal and chl-a frontal zones in the Celtic Sea, identified using the same front frequency indices, have been found to be significant foraging features for breeding northern gannets Morus bassanus (Scales et al. 2014a). When considered together, these results suggest that persistent mesoscale frontal zones in UK shelf seas may have significant cross-taxa ecological importance, providing spatio-temporally predictable foraging opportunities for both planktivorous and piscivorous marine vertebrates.

Associations with Contemporaneous Thermal and Chl-a Fronts

Basking sharks were found to associate strongly with productive regions of the study area, indicating that the propagation of surface-foraging opportunities is tightly coupled with bottom-up oceanographic forcing. Our analysis also reveals that over timescales of weeks to months, sharks associated with thermal and chl-a fronts within these productive areas. Time-matched front metrics were significant predictors of shark presence at the surface. Tracking locations were more likely to be found in close proximity to thermal and chl-a fronts, at higher thermal gradient densities and in association with more persistent thermal fronts than pseudo-absences derived from random walk simulations. These findings are in concordance with those of Curtis et al. (2014), and with our a priori assumption that foraging behaviour of these planktivores is closely tied to low trophic-level enhancement. Comparable associations with thermal fronts in pelagic waters have been documented in other sharks and large teleosts, including the blue shark Prionace glauca (Queiroz et al. 2012), ocean sunfish Mola mola (Sims & Southall 2002), bluefin Thunnus thynnus (Schick, Goldstein & Lutcavage 2004), albacore Thunnus alalunga and skipjack Katsuwonus pelamis tunas (Fiedler & Bernard 1987) and swordfish Xiphias gladius (Podestá, Browder & Hoey 1993; Seki et al. 2002) in differing oceanographic regions, suggesting that thermal fronts could have multitaxon ecological importance for pelagic predators.

Furthermore, basking shark presence was more likely to be associated with lower SSTs, indicating that fine-scale upwelling and vertical mixing are likely to influence the propagation of profitable foraging opportunities. Upwelling fronts are sites of strong biophysical coupling, along which nutrient retention and vertical mixing increase primary productivity and attract grazers such as the calanoid prey of basking sharks (Smith et al. 1986; Franks 1992a; Sims & Quayle 1998; Shanks et al. 2000).

Through systematically varying the threshold used for the detection of thermal fronts, our analysis has revealed that cross-frontal temperature difference is likely to be an important influence on foraging decisions. Regression coefficients and proportion of deviance explained across the model iterations per threshold indicate that stronger (1·0 °C cross-frontal step) fronts have more influence over the probability of shark presence than thermal fronts with a weaker cross-frontal temperature step. In addition, the effect of the gradient density gdens metric indicates that sharks are more likely to associate with stronger fronts. While part of this effect may be related to the spatial element of this study, in that stronger fronts are less numerous and so less likely to be encountered by random walk simulations, this nevertheless indicates that tracked sharks were found in closer proximity to these strong fronts than could be expected by chance. These findings highlight the importance of the choice of front detection threshold in studies investigating species–habitat relationships. The influence of relative sizes of fronts detected has not been explicitly considered here owing to methodological considerations, but may be an interesting subject for future research.

The magnitude of cross-frontal temperature difference is likely linked to persistence and the degree of bioaggregation occurring at a front, owing to the spatial and temporal lags inherent in biophysical coupling mechanisms (Le Fevre 1986). Stronger fronts are more likely to persist through time and also potentially more likely to attract foraging sharks. The mechanisms through which basking sharks detect and respond to environmental clues associated with biophysical coupling at fronts are not yet well understood, but frontogenesis and front propagation are likely to induce the development of discernible environmental clues (e.g. surface and subsurface flow patterns, tidal slicks and streams, accumulation of biota; Franks 1992b). These cues are likely to be more pronounced in the vicinity of stronger, more persistent fronts.

Modelling the influence of contemporaneous fronts on habitat selection has revealed that spatio-temporal persistence of thermal fronts is an important aspect of their attractiveness as surface-foraging hotspots. Thermal fronts in shelf seas around Great Britain and Ireland form primarily as a result of interaction between tidal processes, seasonal stratification and bathymetric influence (Pingree & Griffiths 1978; Simpson & Sharples 2012). As a result, fronts range from ephemeral, only manifesting at certain stages of the tidal cycle, to quasi-stationary and seasonally persistent (Belkin, Cornillon & Sherman 2009; Simpson & Sharples 2012).

Persistent fronts are more likely to be sites of bioaggregation (Bakun 2006), and hence more likely to attract foraging marine vertebrates, than ephemeral features. While gannets in the Celtic Sea appear to target foraging effort within seasonally persistent frontal zones, responses to contemporaneous fronts are highly variable (Scales et al. 2014a). We here provide evidence that basking sharks may associate with contemporaneous fronts more actively than these piscivorous birds, and while persistence evidently has an influence, sharks may also associate with more ephemeral features. We can surmise that aggregation of the sharks’ preferred zooplankton prey does not involve the same spatial and temporal lags that would be required for bioaggregation to propagate through the food chain from plankton to pelagic fish populations and, in turn, to their predators. This work highlights the importance of persistence, and spatio-temporal predictability, of fronts when considering their value as habitats for marine predators.

Technical Limitations

While this study enhances understanding of associations between basking sharks and fronts in the north-east Atlantic, it is not of course without limitations. Using archival tracking technologies based on light-level geolocation has intrinsic limitations, owing to the low level of spatial accuracy of location estimates. However, we have propagated this uncertainty through modelling by repeatedly resampling potential presence locations from within an experimentally-derived radius of error around each geolocation estimate, and randomly resampling from this presence data set before fitting each model iteration. The future use of more accurate tracking technologies, such as fast-acquisition GPS systems (e.g. Fastloc^™-GPS; Wildtrack Telemetry Systems Ltd., Leeds, UK), will enable finer-scale investigations into the drivers of habitat preference in this species and other pelagic marine vertebrates (e.g. Sims et al. 2009). The use of GPS-based tracking with composite front mapping or similar techniques would be a logical follow-up to the results presented here.

Moreover, our study has been restricted to the analysis of movements of only a few individuals (n = 7) over part of one year of their life cycle, so we are hesitant to extrapolate findings to the population level. Many aspects of the life cycle of the basking shark remain unknown, including the size of the population using shelf seas of the north-east Atlantic, and longer range migratory behaviour (Sims 2008). We cannot ascertain whether fronts are significant habitat features for basking sharks throughout the annual cycle or throughout their range. In the north-west Atlantic, tracked basking sharks move from higher latitudes in summer to equatorial regions in winter (Skomal et al. 2009), but in the north-east Atlantic other tracking work has revealed that the shelf-break system, a region of frequent and intense surface frontal activity, may represent an important over-wintering habitat (Sims et al. 2003).

Results presented here indicate that sharks also associate with thermal and chl-a fronts manifesting in coastal waters of the region in summer, when sharks frequently feed at the surface and occasionally dive to the sea bottom (Sims et al. 2005), and so are at their most vulnerable to deleterious anthropogenic interactions [e.g. fisheries bycatch; development of Marine Renewable Energy Installations (MREI); impacts of maritime leisure]. Composite front mapping is useful in identifying key habitats and potential regions of overlap with anthropogenic pressures within the Exclusive Economic Zones (EEZ) of nations, and so could be of value in marine spatial planning and the formulation of management initiatives for species of conservation concern (Miller & Christodoulou 2014; Scales et al. 2014b).

Although oceanographic front metrics derived from composite front mapping have proven useful in this context, the technique has some constraints that must be taken into account. Along with all marine remote sensing applications, only the surface profile of complex three-dimensional oceanographic processes can be detected. However, surface frontal activity can be a useful indicator of subsurface biophysical processes that influence prey availability (Le Fevre 1986; Genin et al. 2005). Moreover, this study focuses on basking sharks that spend long periods surface-feeding, which may be more closely associated with surface frontal activity than other deep-diving marine vertebrates (e.g. northern elephant seal Mirounga angustirostris; Robinson et al. 2012). In addition, the spatial resolution of SST and chl-a imagery used to derive the front indices is limited by the satellite-based sensors. Here, we use LAC to obtain 1·1 km resolution products, but we cannot detect finer-scale oceanographic influence on shark movements. The issue of spatial resolution has an impact on the algorithm's ability to detect fine-scale tidal mixing fronts occurring near to the coastline, which have been identified as potentially significant features for marine vertebrates utilizing the nearshore coastal zone (e.g. Jones et al. 2014). However, front metrics used here are appropriate for oceanographic contextualization of animal movements occurring across pelagic seascapes over timescales of days–weeks–months, complementing the recent proliferation of data obtained through biologging.