Ross seals were captured, instrumented and sampled in the King Haakon VII Sea off Queen Maud Land (6954′S–7220′S and 200′W–1746′W) and in two cases in the marginal sea-ice zone, north of the Lazarev Sea around ∼59°S and 5–21°E in spring 2019 (Supplementary Table 1). We combined tracking data of Ross seals collected within this region during South African National Antarctic Expedition S55 (December 2015–February 2016), the German Antarctic Expedition PS111 (January 2018–March 2018), and the South African Southern Ocean Seasonal Experiment a.k.a. SCALE (October–November 2019), with published data from Norwegian Antarctic Research Expedition (NARE 2000/01) from 2000 to 2001 (Blix and Nordøy, 2007). For brevity, hereafter we refer to the King Haakon VII Sea, Lazarev Sea, Weddell Sea (predicted to in habitat models–see below), and the rest of the area bounded by 40°S to 80°S; 80°W to 80°E as “the Weddell Sea and adjoining areas.”

Ross seals were captured and physically restrained on pack-ice floes in austral summers of 2016 (n = 11), 2018 (n = 2), and spring of 2019 (n = 2) as the animals were encountered along the cruise track of the MV SA Agulhas II (2016, 2019) and RV Polarstern (2018, Alfred-Wegener-Institut Helmholtz-Zentrum für Polar- und Meeresforschung, 2017). An A-frame net, consisting of two stainless steel poles, hinged at the front end, with a nylon net in between, were used to capture the seals. After restraining the seal, a small hole was cut into the net where the top of the seal’s head was located, to place the satellite transmitter. Quick-setting Araldite Epoxy resin (AW2101/HW2951) was used to glue the transmitter to the seals’ heads. The capture and tagging procedure followed Arcalís-Planas et al. (2015) and animals were restrained for a maximum of 1 h. We determined the sex of the animals, and took standard length and girth measurements (Bonner and Laws, 1993). The age of individuals is unknown. The standard measures are not exactly comparable to the NARE 2000/01 data where animals were anaesthetized and therefore fully relaxed and extended (Blix and Nordøy, 2007), but are comparable to those of Arcalís-Planas et al. (2015). During the 2015/2016 summer voyage (hereafter S55) six SPOT6 and five Splash-series tags (Wildlife Computers, Redmond, WA, United States) Argos-linked (CLS, Toulouse, France) satellite trackers were deployed. During summer 2018 (hereafter PS111) two SPLASH10-309A tags, and during SCALE in spring 2019, two SPOT6 tags, were deployed. Satellite transmissions from the satellite tags were not duty cycled and were allowed data to be transmitted every day for all hours of the day, and were restricted to 500 accumulated transmissions per day. Transmissions were paused when the animal was hauled out for longer than 12 h, but allowed to transmit again if the haul out lasts longer than 8 days. Transmissions relayed data that were collected the previous 2 days. Methods on the capturing, immobilisation and deployment of satellite trackers during the Norwegian Ross seal study from 2001 to 2002 are in Blix and Nordøy (2007). The Norwegian dataset consists of Argos-collected satellite tracking data from 10 individuals fitted with SDR-T16 trackers (Wildlife Computers, Redmond, WA, United States).

Location data collected through the ARGOS satellite system contain intrinsic errors. To account for this, we fitted a two-state, behaviourally switching, state-space model to individual tracks (Jonsen et al., 2005; Jonsen, 2016). This procedure filtered erroneous location estimates and provided interpolated tracks with estimated locations at 3 h time intervals. This time step was based on a combination of the median number of Argos location points per day (∼30 points per day) and following logic on available temporal resolution of the environmental covariates (i.e., it does not make sense to have 30 points per day compared with, for example, only two values per day of sea-surface temperatures). Bayesian State-space models were fitted using Markov chain Monte Carlo in “rjags” (Plummer, 2016), via the “bsam” library (Jonsen et al., 2005; Jonsen, 2016) implemented in R (R Core Team, 2020). Two Markov chains were run in parallel, each of 55,000 iterations, only using every 50th value, while the first 10,000 values (i.e., burn-in) were excluded. Diagnostic plots were used to assess converging and appropriate mixing of the two Markov chains (Jonsen et al., 2013). Apart from filtering and interpolating the tracking data, state-space models also identify two hidden states in the movement data and classify location estimates into either “travelling” and “area-restricted search” behaviour, which is based on the animal’s swimming speed and turning angle (Jonsen et al., 2013). Animals travelling between points are assumed to swim faster and in a straight line to cover a large area in low detail, whereas animals searching, hunting or foraging, are expected to swim slower and turn more often to cover a small area in detail, making area-restricted search a proxy for prey searching or hunting behaviour (e.g., Fauchald and Tveraa, 2003; Patterson et al., 2008).

We calculated 50% kernel utilisation distributions (i.e., core habitat use; Worton, 1989) for each individual split between the austral seasons (Summer: October–March; Winter: April–September) using the R library “adehabitatHR” (Calenge, 2006). H-values were selected using the ad hoc method (Silverman, 1986). Kernel densities were plotted for illustrative purposes and data exploration.

We used 16 remotely sensed environmental covariates to describe the habitat use of Ross seals (Table 1 and Supplementary Table 2). Environmental variables chosen for the habitat modelling are known to affect marine predator foraging behaviour and are commonly used to model their habitat use (Raymond et al., 2015; Reisinger et al., 2018; Wege et al., 2019; Hindell et al., 2020). Variables were extracted from the Australian Antarctic Data Centre using the R library “raadtools” (Sumner, 2015) and processed further if needed using R library “raster” (Hijmans, 2016). We matched the date and time of each seal location to the nearest environmental data in space and time for the dynamic environmental covariates (Supplementary Table 2). To avoid the inherent collinearity among environmental variables, we filtered for the most informative set by calculating variance inflation factors for the entire Weddell Sea region and immediately adjoining waters to the east and west (i.e., 40°S to 80°S; 80°W to 80°E) using the R library “fmsb” (Nakazawa, 2018). Variables with a variance inflation factor larger than 10 were excluded, because this is a good indication of strong collinearity (Nakazawa, 2018).

Species distribution models were created using a supervised machine learning technique, boosted regression trees. We separated the summer and winter portions (see Kernel density estimation above) of the tracks to create a species distribution model for each season. Boosted regression trees were chosen because they are known to perform well in predicting species’ distributions (Reisinger et al., 2018; Wege et al., 2019; Hindell et al., 2020). The state-space modelled “area restricted search” or “travelling” estimates were used as the binary response variable inherently turning these models into classification tree ensembles with a Bernoulli distribution. Given that there are not enough tracks in different years, we did not consider any annual variations. We trained models using the “caret” library in R, which employs functions from the R library “gbm” (Ridgeway, 2015; Kuhn et al., 2019). For summer tracking data we held out 30% of the data (test data, n_summer = 3,241; n_winter = 4,784) with an equal number of area-restricted search and travelling points in each test data set. The remaining training data (n_summer = 11,669; n_winter = 11,163) was unbalanced between the number of “area-restricted search” and “travelling” points (area-restricted search_summer = 7,835 vs. travelling_summer = 3,834; area-restricted search_winter = 11,524 vs. travelling_winter = 2,384). We made use of a bootstrapped modelling approach. The bootstraps involved running 500 boosted regression tree models independently, using random sub-sample of 2000 area-restricted search and travelling points, respectively, which was repeated for summer and winter data. This method balanced the data sets, which reduces overfitting of models and the amount of spatial-autocorrelation that is inherent in tracking data (Hijmans, 2012). Each bootstrap model was tuned respectively using the “tuneGrid” function in the “caret” library, through compiling a range of candidate models and choosing the best model and optimal set of hyperparameters based on the lowest Area Under the Curve (AUC) value from the Random Operator Curve (ROC). Candidate models for each bootstrap made use of a 10-fold cross-validation approach. We calculated the out-of-bag AUC value as the goodness-of-fit measure for each of the bootstrapped models using the 30% hold-out test data.

To generate a prediction map of potential foraging habitat of Ross seals we calculated a seasonal mean value per grid square at the same spatiotemporal resolution applied to each environmental variable used in each of the model bootstraps (Reisinger et al., 2018; Wege et al., 2020). We predicted potential foraging habitat of Ross seals for the entire Weddell and adjacent seas and further to 80° easterly and westerly longitudes and from the coastline to 50°S Latitude and beyond for each bootstrap for summer and winter data, respectively using the “predict” function in the R library “raster” (Hijmans, 2016). Using the out-of-bag AUC value of each bootstrap as the weight of each bootstrap, we calculated a weighted mean average probability of area-restricted search behaviour of the 500 bootstraps using the “weighted mean” function in R library “raster” (Hijmans, 2016). The relative influence (%) of each environmental predictor variable was calculated for each bootstrap for the summer and winter ensembles and averaged across bootstraps. Partial dependence plots of the predictor variables, for each of the bootstraps were created in R library “pdp” (Greenwell, 2017) and averaged to create a mean ± standard deviation partial dependence plot.

Instrumented Ross seals travelled away from the pack-ice and spend most of their time in the open ocean (Blix and Nordøy, 2007; Arcalís-Planas et al., 2015, this study) and traversed the Atlantic sector of the Southern Ocean within a very narrow longitudinal band four times a year (Figure 1 and Supplementary Animations). The likelihood of area-restricted search behaviour in summer and winter covaried with ice concentration variability (Figures 5, 6), which serves as a proxy for sea-ice edge zones and accessibility of ice-covered areas (Wege et al., 2020). The sea-ice edge zones are productive areas due to the input of nutrients in the water column as the sea ice melts, which promote phytoplankton blooms which in turn trigger grazers and other mid-level and higher-order predators to aggregate to forage (Nicol, 2006; Arrigo et al., 2008; Riekkola et al., 2019). The thick multi-year pack-ice present in the Weddell Sea and an increased swimming distance to the open ocean likely makes the southern Weddell Sea an unsuitable habitat for Ross seals, which would explain the low habitat suitability within the Weddell Sea embayment year-round (Figure 3). Bester et al. (2020) reviewed all observational data of Ross seals and found that during their annual moult (late summer, early autumn), they are absent within the inner reaches of the Weddell Sea south of 73°S and west of 30°W. Our habitat model showed that within the Weddell Sea embayment, a small area is still likely foraging habitat (Figure 3). Bester et al. (2020) used only observational data (i.e., Ross seals hauled out to moult/rest), while in contrast our study predicts potential foraging habitat based on environmental variables. It is also important to consider that a habitat model only predicts on the variables given to it, and that there are potentially unknown or currently immeasurable variables that dictate Ross seal foraging habitat.

In summer, the foraging behaviour of Ross seals can be split into two stages: a period close to the narrow continental shelf off Queen Maud Land during the annual moult and open ocean foraging between 55° and 65°S, south of the Polar Front. Open ocean foraging happens either post-moult when seals travel north for the winter, post-breeding (November–December) when breeding individuals return from the ice, and for non-breeders that remained at their northern range throughout the summer (Figures 2, 3). This explains why Ross seal foraging probability increases as the waters become deeper (negatively correlated) and preferred to forage over abyssal water (>4000 m deep; Figure 5). However, there is a peak in the probability of area-restricted search behaviour between 1000 and 1500 m depth (Figure 5), which is considered the continental slope and where Ross seals forage during their annual moult (Southwell, 2005; Blix and Nordøy, 2007; Arcalís-Planas et al., 2015).

Stomach content analyses from earlier investigations suggested that Ross seals sampled in January off Queen Maud Land predominantly prey on Antarctic silverfish and squid (Skinner and Klages, 1994). Antarctic silverfish are pelagic and only occur close to the Antarctic continent, on the continental shelf, or on the shallow slope areas in cold water temperatures ranging between (−1.75 to −2°C) (DeWitt, 1970; Hubold, 1984; DeWitt et al., 1990; Ekau, 1990; Eastman, 1993; White and Piatkowski, 1993; Kellerman, 1996; Trunov, 2001; Donnelly et al., 2004; Davis et al., 2017; La Mesa et al., 2019), and higher temperatures of up to 4°C around islands of the Scotia Arc (Mintenbeck and Torres, 2017). Ross seals are likely only forage on Antarctic silverfish during their annual moult, while over the continental shelf and slope. Based on our understanding of Ross seal movements outside of the moult, these diet samples collected in January are not a reflection of the year-round diet. This agrees with stable isotope results (Rau et al., 1992; Zhao et al., 2004; Brault et al., 2019) and diving behaviour analyses, which also suggest that Ross seals are open ocean foragers that prey on mesopelagic fish and squid (Blix and Nordøy, 2007). Fisheries in the Southern Ocean primarily focus on krill (Euphausia spp.), toothfish (Dissostichus spp.) and squid (Agnew et al., 2005; Kock et al., 2007; Nicol et al., 2012; Chown and Brooks, 2019), which means Ross seals are not in direct competition with krill and toothfish fisheries. Currently, it is unknown what percentage of the Ross seal diet is comprised of cephalopods, but it is not their dominant prey species (Rau et al., 1992; Skinner and Klages, 1994; Zhao et al., 2004). Potential for conflict with cephalopod fisheries are also unknown, given the scarcity of data on cephalopod fishing in the open ocean around Antarctica (Agnew et al., 2005; Chown and Brooks, 2019).

Siniff et al. (2008) suggested that the pelagic foraging behaviour of Ross seals makes them least susceptible to climate change compared to other Antarctic seals. They may even be considered as potential climate change winners, because the shrinking distance from the continent to the sea-ice edge in theory would require Ross seals not to swim as far to reach potential foraging grounds.

Five out of the six top variables influencing Ross seal probability of area-restricted search behaviour were the same between summer and winter, i.e., sea-surface temperature, distance to the ice edge, ice concentration variability, the mixed layer depth and wind-driven sea-surface height anomalies. Ross seals preferred to spend time in waters between −1 and 2°C (Figures 5, 6) and displayed less area-restricted search behaviour in warm sea-surface temperature anomalies (Supplementary Figures 1, 2) during both seasons. Sea-surface temperature is a dominant predictor in most Antarctic predators’ movement behaviour (Hindell et al., 2020), and coincides with productive areas in the ocean, marine predator abundance (Block et al., 2011), and subsequent prey distribution and availability (Reisinger et al., 2018; Bestley et al., 2020).

During the summer, Ross seal area-restricted search probability decreased with an increased mixed-layer depth, while in winter area-restricted search probability increased with an increase in mixed-layer depth. The mixed-layer depth influences depth of prey aggregations and any fluctuations in the mixed-layer depth will influence how deep seals have to dive (and the amount of energy to expend) to find prey. The difference between the correlations of area-restricted search probability and mixed-layer depth across the seasons is likely linked to location of dives (continental shelf vs. off-shelf) and a general expected increase in mixed-layer depths during winter months (e.g., Sallée et al., 2010). In areas of the ocean covered in sea-ice, melting ice causes freshening of the water column and the mixed-layer depth to shoal (Sallée et al., 2013; Meijers, 2014). However, how the mixed-layer depth is likely to respond to climate change is not reliably known because of the inability of climate models to accurately represent stratification in the winter months, and are often biassed toward shallower mixed layers (Sallée et al., 2013), which means climate models cannot always accurately predict what will happen to mixed-layer depths throughout the Southern Ocean. The interactions between processes that can affect stratification–primarily through momentum and buoyancy fluxes–are complex, and so how the mixed-layer depth will be affected by climate change differs by region and proximity to continental shelf. If, as suggested by climate models, the mixed layer depth shoals, it could potentially be beneficial to diving Ross seals and reduce the amount of energy expended while foraging at depth. In summer, Ross seals displayed more area-restricted search behaviour where ocean currents were weaker (Figure 5), which would also typically increase in strength due to higher wind speeds under future climate change scenarios (Young et al., 2011; Young and Ribal, 2019), and therefore could be disadvantageous to their foraging habitat.

In summer, probability for area-restricted search behaviour and distance from the ice-edge covaried, which coincides with the Ross seals’ northward migration away from the continent after the moult, and post-breeding when breeders also travelled out of the ice again. Post-moult in February, when the Antarctic sea-ice extent is at its minimum, changes in Antarctic sea ice extent would have minimal effect on Ross seal foraging. However, November–December, post-breeding Ross seals return to the open ocean and forage pelagically. Expanding sea-ice extent would mean Ross seals have to swim further north to reach open water post-breeding, which could potentially incur higher energetic costs to individuals. In winter, the relationship between Ross seal area-restricted search behaviour and distance from the ice edge was bi-modal, with peaks in area-restricted search behaviour around 500 and 1500 km away from the ice edge. This likely results from the seasonality in the expanding ice-edge: during the early winter months (April and May) Ross seals are still foraging at sea at their maximal range away from Antarctica and as the ice-edge expands, they move southward toward the ice edge, then following the growing seasonal ice (Animations in the Supplementary Files). If the ice edge recedes, this would mean that Ross seals will have to swim smaller distances to reach the open ocean and to return to the ice edge, which could potentially be beneficial to them and reduce the amount of energy expended swimming to foraging areas (Siniff et al., 2008). In summer, there was more predicted suitable foraging habitat available compared to winter for Ross seals (Figure 3). This is because of the seasonal pack-ice growth in winter and Ross seal’s preference for open ocean foraging, further indicating that receding ice edge under future climate scenarios will reduce the swimming distance to available foraging habitat, and that potentially more foraging habitat will become available. However, in the Weddell Sea and Ross Sea regions of the Southern Ocean the seasonal ice edge was expanding and sea-surface temperatures decreasing [but summer ice extent recently recorded to decrease (Turner et al., 2020)]; while the Amundsen Sea ice losses are increasing along with rising sea-surface temperatures (Zhang, 2007; Bintanja et al., 2013; Meehl et al., 2016). These contrasting ice-dynamics will result in different responses in foraging behaviour by Ross seals to climate change.

Despite the impacts on their energy expenditures (swimming distance and diving depth), their preferred foraging grounds are also influenced through changing sea-surface temperatures, increased wind and current magnitudes, increased wave height, and poleward shifting wind (Constable et al., 2014). How these changes will affect the distribution of prey is critical to understand how Ross seals will respond. Myctophids in the Southern Ocean are predicted to undergo species-specific shifts in their distribution, but overall most species shift toward the poles (Freer et al., 2019), which would require sub-Antarctic predators to travel south from their sub-Antarctic islands toward the ice to forage (Labrousse et al., 2017; Reisinger et al., 2018). Ross seals travel northward (this study), and a southward migration of prey would reduce travel distances and could be advantageous. Cephalopods, which also form a component of the Ross seal diet (Skinner and Klages, 1994), are unlikely to be affected by sea-ice changes, but they are vulnerable to mesoscale fluctuations and ocean acidification that are affecting entire ecosystems (Rodhouse, 2013; Xavier et al., 2018). How fish and squid species will respond to climate change and increased fishing pressures within mesopelagic ecosystems is largely still unknown and most studies are focussed on Antarctic krill (Euphausia superba) (Murphy et al., 2007, 2013).

Crabeater seals (Lobodon carcinophaga) rely on Antarctic krill for 90% of its diet (Hückstädt et al., 2012) while Antarctic fur seals (Arctocephalus gazella) breeding in the south Atlantic also predominantly prey on krill (Reid and Arnould, 1996). Dietary specialisation makes both species susceptible to any variances and shifts in prey distribution and abundance and are not able to buffer climate fluctuations as well as other Antarctic seals and seabirds (Forcada et al., 2008; Hückstädt et al., 2020). Generalist predators are more likely to switch between prey species are likely to fair better than specialist foragers (e.g., Angermeier, 1995; Terraube et al., 2011), like Weddell seals (Leptonychotes weddellii) for example, whose diet varies among many species of fishes, cephalopods, and crustaceans (Burns et al., 1998; Lake et al., 2003). But, due to Weddell seals’ association with fast-ice for breeding, they are likely to lose breeding habitat and increase in competition with emperor penguins (Aptenodytes forsteri) (LaRue et al., 2019; Jenouvrier et al., 2020; Trathan et al., 2020). Here, we only considered Ross seals foraging movements and potential energy expenditure in line with climate shifts. Like southern elephant seals, Ross seals are generalists and are free-roaming pelagic foragers, who can travel great distances to find prey, which is why elephant seals are also suggested to likely benefit from climate change (Costa et al., 2010). Unlike elephant seals who breed on sub-Antarctic islands, Ross seals breed on the pack-ice like crabeater and leopard (Hydrurga leptonyx) seals. Despite the energetic gains from swimming shorter distances to foraging locations, they are most likely to suffer breeding habitat loss, which might increasing breeding substrate competition with other pack-ice breeders, similar to Weddell seals and emperor penguins (LaRue et al., 2019). Southern elephant seals, who also prey on squid and fish (Daneri and Carlini, 2002; Van den Hoff et al., 2003) and can dive to depths of over 2000 m, beyond the diving reach of the Ross seal (Blix and Nordøy, 2007; McIntyre et al., 2010), are likely to compete with Ross seals for prey resources.