Weddell seals were equipped with either Satellite-Relayed Data Loggers (SRDLs) or Conductivity-Temperature-Depth SRDLs (CTD-SRDLs) manufactured by the Sea Mammal Research Unit, University of St Andrews (Boehme et al., 2009). Deployments took place at two sites in East Antarctica: in Prydz Bay near Davis Station (−68.58°, 77.97°, n = 27, years: 2006, 2007, and 2011) and Terre Adélie near Dumont d’Urville Station, (DDU) (−66.66°, 140.00°, n = 22, years: 2006, 2007, 2008, 2009, and 2019) and one site in the Ross Sea near Scott Base (−77.85°, 166.76°, n = 29, years: 2014, 2016, and 2019). Note that the Prydz Bay and the Terre Adélie data sets (excluding 2019) formed the basis of Heerah et al. (2017). Only females were tagged in Prydz Bay and the Ross Sea, in Terre Adélie both sexes were tagged. The tags were all deployed following the annual moult in February. The capture, sedation and tagging procedures followed those described in detail elsewhere (Mellish et al., 2010; Heerah et al., 2013; Shero et al., 2018), with the tags glued directly on the head of the seal. The tags recorded data on a seal’s diving behaviour as well as in situ hydrographic conditions. Summaries of these data were transmitted via the polar-orbiting Argos satellite constellation whenever seals surfaced to breathe (Harcourt et al., 2019). Only seals for which the tag transmitted for longer than 50 days were included in the analyses.

The dive data underwent an initial quality control to remove anomalously deep dives. This was done by calculating a lower edge for the dive duration vs. dive depth scatter plot (using a quantile regression set at 0.05). Dives that fell below this edge were deemed to be too deep for that duration (thereby requiring unrealistic rates of travel) and were flagged for removal. We also removed dives of less than 5 min duration as these were regarded as unlikely to be foraging, and dives deeper than −1000 m as these were deeper than the Antarctic continental shelf (Padman et al., 2010). A location was recorded for each remaining dive by using a state-space-model with a 2 h time step and interpolating from the start time of each dive (Jonsen et al., 2020). Each dive was then allocated a bathymetric depth using GEBCO19 (Gebco Compilation Group., 2020).

A scatter plot of bathymetric depth against the maximum dive depth of each dive indicated that 23.6% of the total 171,201 dives were more than 20 m deeper than the GEBCO19 bathymetry at that point (Supplementary Figure 1). The three sources of error that contribute to this mismatch are (1) errors in dive location estimates, (2) errors in the tag’s depth sensors and (3) errors in the bathymetry. These uncertainties were problematic for allocating individual dives to a behavioural type (e.g., hunting “benthic” vs. “pelagic” prey). Therefore, we calculated the difference in depth between the GEBCO19 bathymetry for those dives that exceeded the bathymetry. We then calculated the 30 percentile of these differences, as this approximates 1 Standard Error of a distribution centred on the red line representing equivalence of a dives depth and estimated bathymetry at that point. This value was ∼100 m. Assuming that these mismatches were symmetrical about the 1:1 line in Supplementary Figure 1, we then categorised any dives that were within 100 m of the GEBCO bathymetry as “benthic” dives and those that were >100 m from the bottom as “pelagic” dives. Dives more than 100 m deeper than the bottom was classified as “unknown” (Supplementary Figure 1).

We compared simple metrics of diving behaviours (maximum depth and dive duration) for benthic and pelagic dives between the three study sites using generalised linear mixed models (GLMMs) with individual seal as a random term and deployment location as the main effect. We also modelled the probability of making a benthic dive relative to the bathymetry at the location of that dive using a logistic GLMM, again with individual seal as a random term. We then used this model to make spatial predictions of where seals were likely to make benthic dives through the entire domain (defined by the extent of seal tracks) at each deployment location.

Utilisation Distributions are a common way of quantifying spatial use by individuals and populations (Laver and Kelly, 2008). We used a raster-based approach to calculate the utilisation distributions (UD), in which we recognised two types of Utilisation Distributions: (i) Individual utilisation distributions (UDi), that is the total area used by an individual seal and (ii) Sample utilisation distributions (UDs) representing the spatial use by the population of seals tagged at each deployment location. We used a raster with a 5 × 5 km grid cell. This resolution corresponds to the estimated accuracy (± 5 km) of state-space-model modelled ARGOS-based elephant seal locations using the same trackers (Jonsen et al., 2020). The UDs analysis required three rasters (Supplementary Figure 2); (i) the density (number) of seals that visited each 5 × 5 km cell, (ii) the mean time spent by each seal in each cell (i.e., the overall mean of the individual UDi raster stack) (iii) the product of these two rasters provides a measure of overall habitat utilisation in terms of seal days per cell. This value is a close analogue of kernel density analysis but has the advantage of not imposing a smoother over the data. With sufficient data this smoothing is unnecessary as we have an empirical, quantitatively assessed, measure of the degree of use per cell. To estimate the degree of representativeness of the sample of seals compared to the actual population of seals at each location, we calculated the cumulative area occupied by increasing numbers of seals (Supplementary Figure 3) (Hindell et al., 2003). Finally, we calculated the percent usage contours by sorting the UDs values in each cell, taking the cumulative sum, and using that to identify user defined contours. While there are many methods for summarising utilisation distributions there is broad agreement on the use of 50% contours (i.e., enclosing 50% of locations) to define the core usage areas and 90% to indicate overall usage (Laver and Kelly, 2008).

First, we calculated the area (km²) of each UDi by counting the number of 5 × 5 km cells visited by each seal and compared these among the three deployment locations using a GLMM which included deployment duration as a random term to account for animals with longer durations potentially having larger UD areas. Seal standard length (m) was included in the model to account for animal size potentially being an important source of individual variation in space use (Hindell et al., 2021). We used length as it was the only morphometric measure common to all studies. We also compared UD_i areas between sexes at DDU, the only site where males were also instrumented, to ascertain if there were differences in habitat use and foraging by sex (Photopoulou et al., 2020).

We fit GLMMs relating the number of days that an individual spent in a 5 × 5 km cell to the bathymetry and ice concentration in that cell. Ice concentration in a cell varied little over the winter months, so we took the mean of the ice concentrations over the study period in each cell. While this may lose some fine-scale detail, the Antarctic continental shelf is fully covered by sea-ice by April (Eayrs et al., 2019) and so inconsequential (study lasts March–October). Prior to analysis, outliers (1 and 99% percentiles) for both the response and predictor variables were trimmed from the data set. Predictor variables were also centred and scaled to unit variance to facilitate efficient optimisation of the GLMMs. The distributions of each variable were checked for normality and the response variable (seal days per cell) was subsequently log-transformed. Inter-correlation between the predictor variables were checked with cross-correlation analyses, comparing Pearson Correlation coefficients. The full model GLMMs (with individual seal as a random term) were initially run with and without a spatial autocorrelation term to test its effect on model performance. We found that the spatial autocorrelation term did not affect model performance and so all subsequent analyses were run excluding it. Comparison of log-likelihoods and corrected Akaike Information Criterion (AICc) indicated that the model performed best without spatial autocorrelation, so this was not used in the final model assessment. Overall model fit was assessed with the conditional r-squared which provides the variance explained by the entire model, i.e., both fixed effects and random effects. Finally, we ranked the full suite of 19 models by AICc to identify the best performing model (Burnham and Anderson, 2002, 2004).

We assessed qualitatively the physical characteristics of the seals’ habitats at the population level by contrasting the distribution of bathymetry and sea-ice concentrations within three locations, (i) the core UD_s (50%) the overall UD_s (90%) and the regions potentially available to the seals. The UD_s derived from observed locations only provide information about where animals occur, not about where they could have gone. To estimate potentially available geographic space, we simulated sets of tracks for each observed track using the availability R package¹ simulating 10 tracks per seal. This yielded simulated tracks with movement characteristics (distributions of step length and turning angle) similar to observed tracks, but tracks were random and independent of environmental effects. These simulated tracks provide an estimate of the geographic space that each animal could have occupied (given its movement characteristics and track length) if it had no habitat preferences (Raymond et al., 2014; Hindell et al., 2020). We compared the ice concentration on July 30 (averaged for all years of the study) as this was the time when ice extent was nearing its maximum, and when most tags were still transmitting.

Analyses were conducted in R 4.0.2 (R Core Team, 2020). FoieGras was used to filter the seal tracks (Jonsen et al., 2020). Descriptive statistics presented as mean (X) and standard deviation (±s.d) unless indicated. All models were Generalised Linear Mixed Effects Models in the R package LME4. The models were assessed by ranking candidate models using AICc, and evidence ratios calculated for each model comparison to provide a2n indication of the importance relative to the top model (Burnham and Anderson, 2004).

Tags transmitted for 150.1 ± 72 days (mean ± s.d. Table 1 and Supplementary Figure 3) and this did not vary between sites (Supplementary Table 1a). Seals from all three deployment locations occurred almost exclusively on the Antarctic continental shelf (seafloor >−1000 m) with seals staying within 200 km of the central point of their distribution (Figure 1). This central point was not where they were tagged, because they were tagged when aggregated to moult. Post-moult, the animals moved to other areas as the sea-ice grew. The seals then remained relatively local with individuals moving on average 138.2 ± 117.8 km from that central point. This differed among the three deployment locations; seals tagged in the Ross Sea travelled further than seals in Prydz Bay (mean maximum distance = 178.0 ± 147.3 km compared to Prydz Bay mean max = 151.0 ± 89.1 km) and Terre Adélie (mean max = 70.5 ± 70.4 km) (Table 1).

The best model relating the number of seal days per 5 × 5 km pixel included location, bathymetry, ice concentration and the interactions of deployment location with bathymetry and bathymetry with ice concentration, an AICc value of 12538.1 and an AICc weight of 0.996 (Supplementary Table 1f). When considering main effects, bathymetry was best of the single term models with an AICc of 12670.7, followed by ice concentration (12776.0) and then location (12784.3). The null model had an AICc of 12794.9. Intensity of use of a cell was negatively related to bathymetry with the areas of most intensive use in shallowest waters, and this was most pronounced in Terre Adélie (Figure 2C). However, when considering the interaction terms, the intensity of use was slightly higher in high ice concentration over deep water (Figure 2C).

Our sample sizes, while not large enough to capture the estimated maximum area used by a theoretical population of seals (Supplementary Figure 4), did cover more than 85% of the estimated asymptote. In Prydz Bay, the Gompertz model fitted to the saturation curve had an asymptote of 58335 km², of which our 26 seals occupied (91.8%). In Terre Adélie the 22 seals occupied 86.7% of the estimated 21185 km², and in the Ross Sea the 28 seals occupied 91.0% of the estimated 66724 km².

The core (50%) population UD_s were in shallow waters within 200 km of the central pixel (Figure 3) at all locations. In Prydz Bay, the core UD_s extended northward along the coast, largely contained within the −500 m bathymetric contour. The seals in Terre Adélie had the smallest core UD_s mostly within 50 km of the central point and also within the −500 m bathymetric contour. The core UD_s for seals in the Ross Sea was concentrated on the western edge of McMurdo Sound extending 50 km north of Ross Island. In Prydz Bay the core 50% UD_s with an area of 1700 km² occupied 12.6% of the overall 90% UD_s of the sample of seals at that site. In Terre Adélie the core UD (200 km²) occupied 10% of the overall 90% UD_s, and in the Ross Sea (core = 1650 km²) 11.1% of the 90% UD_s.

The bathymetry available to the seals differed from those in the overall and core UD_s, indicating a strong preference for shelf waters, with randomised tracks venturing off continental shelf break, with mean values deeper than −1000 m for seals from Prydz Bay and Terre Adélie (Figure 4 and Supplementary Table 2). Seals from the Ross Sea showed a different pattern, with available habitat almost exclusively contained to shelf. The bathymetry within the core UD_s also differed from the overall (90%) UD_s (Figure 4 and Supplementary Table 2). At all three locations, the core areas were in considerably shallower water than the overall area (Supplementary Table 2), reflecting the coastal positions of the core UD_s. In the Ross Sea overall UD_s had the deepest bathymetry (mean = −608 ± 147 m) followed by Prydz Bay (−370 ± 112 m) and then Terre Adélie (−262 ± 118 m) reflecting the topographic characteristics of the locations. Ice characteristics varied between available, core and overall areas (Figure 4 and Supplementary Table 2). Ice concentrations were consistently higher within regions available to the seals (Figure 4) compared to regions they actually used. Further, ice concentration was slightly lower in core UD_s than overall UD_s, although the differences were relatively minor, being largest in the Ross Sea (85.0 ± 3.7 [90%] compared to 79.4 ± 2.2 [50%]).

The area of shallow water where benthic foraging was most likely [i.e., ≥−250 m (Figure 2A)] constituted 85% of the core UD in Terre Adélie, 85% in Prydz Bay and only 36% in the Ross Sea (Table 2).

Weddell seal foraging occurs in a mix of habitat types (benthic, epipelagic, and mesopelagic) and they have a catholic diet taking benthic fish, crustaceans and cephalopods as well as mesopelagic prey such as Pleurogramma spp. (Supplementary Table 3). The likelihood of benthic foraging increased in shallower water. To reach the benthos in these areas, seals traverse the entire water column rather than stopping to feed in the pelagic zone, either because the mesopelagic prey are not there or because the seals get a greater energetic return from benthic prey. Harcourt et al. (2007) found that males that dived deeply during the breeding season in McMurdo Sound lost weight at a slower rate than those remaining shallow. This suggests that deeper prey resources are especially profitable or at least predictable (Photopoulou et al., 2020). Making diet inference based simply on predator dive behaviour is potentially confounded by the knowledge that some key prey species such as silverfish and toothfish can occupy multiple habitats (O’Driscoll et al., 2011). Feeding studies and acoustic observations of silverfish are consistent with a predominantly pelagic niche (Pinkerton, 2017; O’Driscoll et al., 2018), but adult silverfish are sometimes found close to the bottom over the Ross Sea shelf (O’Driscoll et al., 2011). In contrast, toothfish over the Ross Sea shelf are likely to be predominantly demersal (Pinkerton et al., 2016), but pelagically feeding in McMurdo Sound (Ainley et al., 2013).

As seals venture into deeper water, the benefits of benthic foraging are likely to decline due to the energetic trade-offs. These are associated with deeper diving, such as reduced time available in the foraging zone, greater cost of transport to and from deeper depths (Thompson and Fedak, 2001; Williams et al., 2004), physiologic costs associated with exceeding the aerobic dive limit or even potentially cardiac anomalies (Williams et al., 2015). This may explain why the seals in all three regions used predominantly shallower habitats in accordance with their relative availability.

In this study there was a high degree of inter-individual variation in winter foraging behaviour. Some individuals were notably sedentary, did not move far from tagging location and had small home ranges. Others had large home ranges, travelled several hundred kilometres, and made more pelagic dives. These differences may be due to divergent behavioural traits among individual seals (Toscano et al., 2016; Troxell-Smith and Mella, 2017). Individual behavioural traits under-pins differences in foraging behaviour in other species (Patrick et al., 2014; Toscano et al., 2016; Krüger et al., 2019; DiNuzzo and Griffen, 2020; Steinhoff et al., 2020), and distinct, stable foraging strategies have been identified in many pinnipeds [e.g., Australian sea lions, Neophoca cinerea (Lowther et al., 2011); Galapagos sea lions, Zalophus wollebaeki (Villegas-Amtmann et al., 2008)], consistent with individual traits (Schwarz et al., 2021). In dynamic environments, multiple foraging strategies within a population may reduce resource competition (Lewis et al., 2006; Baylis et al., 2015). We found no relationship between home range area and seal size. In northern elephant seals, Mirounga angustirostris, body condition influences foraging behaviour, with animals in poor condition adopting riskier foraging strategies (Beltran et al., 2021) but, post-tagging we have no information on individual condition on our animals. Development of in situ proxies of condition (e.g., buoyancy) in Weddell seals similar to those well established in elephant seals (Biuw et al., 2003) could help to resolve whether these differences result from intrinsic strategies or arise from variance in foraging success. Finally, pre- and post-tagging reproductive status may both be important. The majority of these animals were females and at the time of tagging some would have just weaned their pup from the preceding breeding season, while others were not reproductive that season. Animals that skip breeding have different energetic demands as they have avoided the high cost incurred by birthing and lactation, where females lose about 30% of their body mass (Wheatley et al., 2006; Shero et al., 2015). Females that skipped breeding are in much better condition and so may have the luxury of adopting a safer foraging strategy as with fat northern elephant seals (Beltran et al., 2021). Conversely, not all females are pregnant each year. Pregnancy incurs increased energetic demands which translates to increased winter foraging effort (Shero et al., 2018), and this may well influence space use.

Our results differ from studies of Weddell seal distributions in the Weddell Sea (Boehme et al., 2016; Nachtsheim et al., 2019; Photopoulou et al., 2020; Labrousse et al., 2021), where tracked individuals showed diverse and wide-ranging movements including moving off the Antarctic continental shelf. In those studies, seals were often captured from ships some distance from the coast and later in the winter. In our study, all seals were captured just after their annual moult (∼2–3 months after the breeding season) at or near their breeding sites on coastal fast ice. We hypothesise that the seals tagged offshore in the Weddell Sea may contain a higher proportion of those individuals that are predisposed to make longer excursions, and so represent a different sample from those tagged near their breeding areas. Supportive evidence arises from Nachtsheim et al. (2019) who were forced to tag four of their six animals in a sheltered inlet due to bad weather. They found that the four Weddell seals they tagged on the fast ice stayed close to their colony, while the two tagged on the pack ice ventured further afield.

Bathymetry and ice concentration were important determinants of Weddell seal distribution. Sea-ice concentration on the continental shelf in east Antarctica is dynamic both within and among years (Massom et al., 2013), but in most years ice concentration reaches a maximum in late winter, around August (Eayrs et al., 2019), while still exhibiting considerable spatial variation due to recurrent coastal polynyas (Arrigo et al., 2015). These seals did not concentrate in polynyas, but instead those individuals that made excursions beyond the core areas moved into areas of relatively high ice concentrations (80% plus). Inhabiting high sea-ice concentration areas may improve access to prey or reduce the risk of attack by predators such as killer whales (Lauriano et al., 2020). Similarly, sedentary individuals remaining in the fast ice may reduce predation risk.

Antarctica and the surrounding Southern Ocean, like many regions across the globe, are changing in response to changes in climate (Rogers et al., 2020) and anthropogenic activities such as increased exploitation (e.g., fisheries), and these can profoundly affect endemic biota (Bestley et al., 2020). To quantify how these broad scale changes affect individuals and animal populations requires information on animal distribution and behaviour, and the structure and dynamics of the physical environment (Hindell et al., 2020). Quantifying movement and habitat use by predators for conservation and management strategies is particularly pertinent given the recent designation of the Ross Sea region Marine Protected Area (MPA), and the ongoing debate for new MPAs in east Antarctica and elsewhere in the Southern Ocean (Hays et al., 2019; Hindell et al., 2020). It is especially important in the management context to develop policies that protect the endemic biota over the long-term while regulating the human behaviours like fishing within the broader framework of a changing climate, over which it is much harder to effect change (Hays et al., 2019; Hindell et al., 2020). In the Southern Ocean, animal conservation and fisheries regulation is generally managed by the Commission for the Conservation of Antarctic Marine Living Resources (CCAMLR) under an ecosystem-based management approach which includes establishing a network of MPAs in the Southern Ocean. Information on where, when and how animals use their habitats is critical for the design and evaluation of these MPAs.