Potential for redistribution of post- moult habitat for Eudyptes penguins in the Southern Ocean under future climate conditions

Anthropogenic climate change is resulting in spatial redistributions of many species. We assessed the potential effects of climate change on an abundant and widely


| INTRODUC TI ON
Anthropogenic climate change is resulting in redistributions of species worldwide with marine species redistributing poleward at a rate six times faster than terrestrial species (Lenoir et al., 2020). The Southern Ocean (defined here as including the southern zones of the Indian, Atlantic and Pacific Oceans, from the Subtropical Front to the Antarctic continent; Deacon, 1982) is one of the most rapidly warming oceans on the planet (Sallée, 2018). Here, abundant populations of marine predators play a key structuring role in marine ecosystems , and act as bio-indicators of the ecosystem state (Hazen et al., 2019). Many marine predators may be threatened by the rapid changes taking place in the Southern Ocean (Barbraud et al., 2011;Constable et al., 2014;Le Bohec et al., 2008;Trathan et al., 2007). Despite this, the first Marine Ecosystem Assessment for the Southern Ocean (MEASO) highlighted the lack of knowledge on how marine predators will be redistributed across their ranges ; but see Cristofari et al., 2018;Reisinger et al., 2022). The MEASO project aims to collate research and assess trends on Southern Ocean ecosystems under climate change (with the aim of enabling policy makers to achieve consensus in adapting their management strategies to ecosystem change). Conservation and management of these animals relies on understanding their distributions, how these relate to the bio-physical environment (Reisinger et al., 2018); and how these distributions may alter in a changing marine environment (Pecl et al., 2017). Predicting how distributions of species may change is critical in predicting their potential for adapting to climate-related environmental change (Jenouvrier et al., 2014;Rose et al., 2010).
Marine predators, such as penguins, are limited in their capacity to respond to rapid changes in their environment and the resulting mismatch in the distribution of foraging habitat and prey Dias et al., 2019;Grémillet & Boulinier, 2009;Morgenthaler et al., 2018). Penguins represent nearly 90% of the avian biomass in the Southern Ocean (Croxall & Lishman, 1987).
However, populations of several species are declining and are predicted to decline further as warming of the ocean continues (Barbraud et al., 2008;Boersma et al., 2020;Jenouvrier et al., 2020;Ropert-Coudert et al., 2019). Penguin species are particularly vulnerable to changes in preferred habitat or prey redistribution distributed group of diving birds, Eudyptes penguins, which are the main avian consumers in the Southern Ocean in terms of biomass consumption. Despite their abundance, several of these species have undergone population declines over the past century, potentially due to changing oceanography and prey availability over the important winter months. We used light-based geolocation tracking data for 485 individuals deployed between 2006 and 2020 across 10 of the major breeding locations for five taxa of Eudyptes penguins. We used boosted regression tree modelling to quantify post-moult habitat preference for southern rockhopper (E. chrysocome), eastern rockhopper (E. filholi), northern rockhopper (E. moseleyi) and macaroni/royal (E. chrysolophus and E. schlegeli) penguins. We then modelled their redistribution under two climate change scenarios, representative concentration pathways RCP4.5 and RCP8.5 (for the end of the century, 2071-2100). As climate forcings differ regionally, we quantified redistribution in the Atlantic, Central Indian, East Indian, West Pacific and East Pacific regions. We found sea surface temperature and sea surface height to be the most important predictors of current habitat for these penguins; physical features that are changing rapidly in the Southern Ocean. Our results indicated that the less severe RCP4.5 would lead to less habitat loss than the more severe RCP8.5. The five taxa of penguin may experience a general poleward redistribution of their preferred habitat, but with contrasting effects in the (i) change in total area of preferred habitat under climate change (ii) according to geographic region and (iii) the species (macaroni/royal vs. rockhopper populations). Our results provide further understanding on the regional impacts and vulnerability of species to climate change.

K E Y W O R D S
climate change, habitat preference models, migration, overwinter, species redistributions, Subantarctic penguins compared to flying birds because they cannot as easily cover large distances to forage. Indeed, metabolic rates of swimming penguins are equivalent to 2.5 times those of soaring albatrosses of a similar body mass (Alexander, 2002). Migrations are energetically costly, and penguins can only perform these post-moult migration round trips of ~10,000 km through high energetic payoffs over this period (Alexander, 2002). Additionally, penguins are restricted by the limited number of breeding/moulting sites accessible to them which reduce their capacity to relocate colonies if resources move (Cristofari et al., 2018).
The penguin genus Eudyptes (crested penguins) consists of eight taxa (seven IUCN recognized species) that breed in locations from the subtropics to the Antarctica peninsula. This penguin assemblage is the main avian consumer in the Southern Ocean in terms of biomass consumption, eating millions of tonnes of prey (including euphausiids, myctophids and cephalopods) annually (Brooke, 2004).
Yet, despite their widespread abundance, several species have undergone population declines over the past century .
They perform two migrations annually: the pre-moult migration occurs immediately after breeding and lasts 1-2 months before their catastrophic moult (when they moult all their feathers at once), and the post-moult migration when they spend 4-6 months overwintering at sea, covering vast distances. During the post-moult migration, the first and last months are particularly important as individuals undergo hyperphagia (periods of intensive foraging) during which they recover condition following their catastrophic moult (throughout which they fast for 4-5 weeks) and gain substantial fat and muscle reserves in preparation for breeding (Green et al., 2005;Thiebot, Cherel, et al., 2011).
However, population declines of marine predators are also mediated by the changing biotic and abiotic oceanographic conditions that influence the distribution and abundance of marine resources (Péron et al., 2012;Trathan et al., 2007). Population changes of Eudyptes penguins, in particular, have been caused by changes in oceanography and prey availability which can impact on nonbreeding survival and on breeding success (Crawford et al., 2008;Crawford & Dyer, 2006;Horswill et al., 2016;Morgenthaler et al., 2018). Therefore, it is important to predict how future conditions under climate change might affect the availability and extent of preferred nonbreeding habitats (which reflects where preferred prey will be distributed) and hence post-moult distributions of Eudyptes penguins.
Here, we investigated the habitat preference and potential redistribution of preferred habitat in response to climate change for five Eudyptes taxa, during their post-moult period. We quantified current physical habitat characteristics of the southern rockhopper (E. chrysocome), eastern rockhopper (E. chrysocome filholi), northern rockhopper (E. moseleyi) and macaroni/royal (E. chrysolophus and E. schlegeli) penguins in the Southern Ocean. In this study, we define the Southern Ocean to include waters ranging from 30 °S to Antarctica following Deacon (1982) and Talley (2011) encompassing the breeding islands and foraging ranges of northern rockhoppers. Thereafter, we identified potential future habitat analogues under two climate change scenarios, Representative Concentration Pathways RCP4.5 (our current projection) and RCP8.5 (worst case scenario). RCP4.5 and RCP8.5 correspond to medium and high radiative forcing, respectively. Radiative forcing is the measure of the amount of downward-directed radiant energy upon the Earth's surface from greenhouse gas emissions, aerosol emissions and solar irradiance (Portner, 2019). RCP4.5 projects global temperatures increasing by 1.1-2.6°C and a mean sea level rise of 0.47 m, by the year 2100 and RCP8.5 projects global temperatures increasing by 3.0-12.6°C and a mean sea level rise of 0.62 m, by the year 2100 (Meinshausen et al., 2011;Pielke Jr. et al., 2022). Our aims were to (1) use multi-species and multi-population post-moult tracking data to infer post-moult habitat preference for five taxa of Eudyptes penguins from nearly all major breeding sites across the Southern Ocean; (2) use two greenhouse gas concentration pathways, RCP4.5 and RCP8.5, to model future potential redistribution of habitat (at the end of the century, 2071-2100); and (3) quantify the magnitude of habitat change per MEASO-defined region (Grant et al., 2021; for both RCP scenarios for each taxon.

| Breeding distributions of study species
Currently, there are two recognized species of rockhopper penguins (IUCN, 2022) but recent genetic studies have found that southern rockhopper penguins should be split into two sister taxa, southern and eastern rockhopper penguins (Banks et al., 2006;Frugone et al., 2021). We therefore separated the rockhopper penguins into three taxa (northern rockhopper penguins, southern rockhopper penguins and eastern rockhopper penguins).
Furthermore, royal penguins are genetically similar to macaroni penguins, despite their clear phenotypic differences Frugone et al., 2019), so we used this opportunity to also predict where royal penguins could forage during their postmoult migration as currently there are no tracking data for this species for this stage of their annual cycle. Therefore, in this study, we included the royal penguin as part of the macaroni penguin group. All species breed on Subantarctic islands ( Figure 1 Table S1 for details).
All statistical analyses were performed using R Statistical Software 4.1.0 (R Core Team, 2021).  Table S2 for details on tracking years and sample sizes). Data are from 10 of the 21 major breeding sites including the Falkland Islands  Table S3 for Data Availability). GLS tags are miniaturized archival data tags with long battery lives, which make them currently the only suitable method to capture a full post-moult migration while minimizing the burden on the animals (Bost et al., 2009). Tags were mounted on the birds' legs (during the breeding season or at the end of the moult, depending on the colony) using purpose-designed bands (e.g. Ratcliffe, Takahashi, et al., 2014). The tags were recovered when the birds returned to their colonies to breed. As well as light level measurements, which can be used with local time to estimate latitude and longitude twice a day, the devices whenever possible also recorded sea surface temperature (SST).

| Tracking data
We analysed 485 full tracks (163, 433 location points in total) from 597 raw light files (raw light files that were faulty or with partial recordings were not included) using the BAStag and SGAT packages in R (S. Wotherspoon et al., 2013b;S. J. Wotherspoon et al., 2013a) with an ice mask around Antarctica. Where tags collected temperature data, a SST prior was used to constrain the birds' location based on Reynolds weekly SST maps (Lisovski et al., 2019;Reynolds et al., 2005).
Prior to statistical analyses, these tracks were then processed using a hierarchical state-space model (SSM) using the bsam package (Jonsen et al., 2020). We did this to estimate locations at regular intervals (12 h) as well as provide movement parameters (trip duration, turning angles, maximum distances from the colony and travel speeds) used to simulate biologically meaningful tracks (see below).

F I G U R E 1
Distribution of breeding sites of the five focal Eudyptes penguin taxa across the Southern Ocean: eastern rockhopper penguins, macaroni/royal penguins, northern rockhopper penguins and southern rockhopper penguins.

| Simulating tracks to estimate the random occurrence likelihood
We used a case-control design for habitat preference modelling of the observed tracking data, where environmental characteristics along the observed tracks were compared to those along a set of simulated tracks (Aarts et al., 2008). This is like a presence-background design in general habitat preference modelling. Simulated tracks represent a set of locations where an animal would move with no habitat preference but with the same movement characteristics (identical dates, trip durations, maximum distances from the colony and travel speeds) of the observed tracks. They represent a set of background location estimates which represent the likelihood of random occurrence and that consider the geographic availability of cells to animals (Raymond et al., 2015). To do this, we simulated 50 tracks for each observed track using movement parameters from the SSM, with the availability package in R (Raymond et al., 2015;Reisinger et al., 2018). This allowed the habitat preference of the animals to be modelled using the environmental characteristics at the locations where the animals were observed and those at locations that they occurred randomly.
The number of simulated tracks were chosen as a balance between adequate coverage of the random use of the marine environment and limiting the dataset size for computation .
Simulated tracks were constrained to an ecologically realistic geographic space that is accessible to these taxa of Eudyptes penguins during their winter period. We have prior knowledge of broadscale habitat use (e.g. the birds do not visit tropical or high latitude areas) so including these would result in the avoidance of these regions dominating the habitat preference response. We therefore latitudinally constrained the simulated tracks to be between 30°S and the northern extent of the ice edge, defined here as <80% ice concentration. This prevented simulated tracks from unrealistically extending into tropical waters to the north or into areas which would be covered by winter sea ice in the south. Simulated tracks were also constrained by a land mask which prevented any tracks being located on land.

| Environmental data
To characterize the bio-physical conditions associated with utilized (observed) and random (simulated) locations, we used a suite of nine environmental covariates (Table 1). Environmental data were remotely sensed, or model estimated, and represented variables known to influence penguin distributions (Péron et al., 2012; or likely to influence the distribution of the prey (Pinkerton et al., 2020). Environmental data from Copernicus Marine Service (www.marine.coper nicus. eu) were used (Table S4). The environmental data associated with the observed and simulated tracks were extracted using the raster package (Hijmans et al., 2015) by matching location and date at a 1° × 1° spatial resolution and daily time scale. Thereafter, to predict preferred habitats at a Southern Ocean wide scale in the habitat preference model step, mean climate data (climatologies) were compiled from the same environmental variables ranging from April to September (austral autumn and winter when these taxa of Eudyptes penguins migrate) spanning the length of time which the tracks covered (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019)(2020)

| Habitat preference models
Boosted regression trees (BRT) are a machine learning algorithm with a high predictive accuracy and abundant flexibility. In this study, we wanted to identify the environmental variables that are important to predict where these are for the global populations now and in the future. The predictive capability of BRT modelling is more robust compared to other models as the technique of boosting to combine large numbers of relatively simple tree models adaptively optimizes the predictive performance (Elith et al., 2006;Elith & Leathwick, 2017). Considering this and the ability of BRT modelling to deal with non-linear relationships and correlated and interacting variables (common in ecological studies; Elith et al., 2006Elith et al., , 2008; we used BRT modelling to model penguin habitat preference as a function of their biophysical environment using the gbm package (Greenwell et al., 2019). BRT modelling was done by species.
We used a learning rate of 0.01 and chose optimal number of trees (minimizing the prediction bias) using a k-fold cross-validation by individual (Elith & Leathwick, 2017). Robust estimates of the number of trees that can be used is important to avoid overfitting as the models need a degree of generality to predict successfully into unsampled areas (Schonlau, 2005). To increase the model accuracy, randomness was included using a bag refraction of 0.5 and tree complexity of 5. A Bernoulli family appropriate to the binary response variable (observed (1) vs. simulated (0)) was stipulated. The model outputs describe the probability of relative habitat preference. The habitat models were evaluated using number of trees, percentage of deviance, R 2 (the proportion of the variance for a dependent variable that's explained by an independent variable) and CV AUC (cross validation for area under the curve). Blocking validations were performed at the colony level (Roberts et al., 2017) to test the predictive performance of the species preferences models on individual colonies, with summary statistics exhibited with confusion matrices.
Although extensive, our tracking dataset lacked data for colonies of rockhopper on Macquarie and Auckland Islands and maca-

roni/royal penguins on Heard and McDonald Islands and Macquarie
Island. To account for these data gaps, we used the habitat preference models to predict preferred habitat for each of the species across their entire ranges, including where no tracking data were available (see Table S1 for full list of breeding location used).

| Accessibility
The data used to fit the habitat preference models (i.e. the observed and simulated tracks) represent habitat only and may include geographic areas that are not accessible to the animals (e.g. outside their foraging range; Aarts et al., 2008;Matthiopoulos, 2003). As penguins return to their breeding colonies at the end of the post-moult migration, it was important to consider the accessibility of a location (given they cannot just keep swimming endlessly). Habitat preference models were therefore constrained using accessibility models (models of areas that the birds could feasibly access) to produce an index that reflected both the habitat preference of a given cell and how accessible that cell is to the birds .
Accessibility was modelled by using observed and simulated tracks as a function of the distance of the cell to the deployment colony.
The number of locations were converted to a binary response (observed and simulated tracks vs. no observed and no simulated location). Then binomial models were fitted with a smooth, monotonic decreasing constraint as we assumed that accessibility decreases the further away from the colony. This additional modelling step was used to estimate which areas of geographic space were accessible to the birds by upweighting cells that were both preferred habitat and accessible and down-weighting cells that were preferred habitat but not accessible. These outputs provide an estimate of the probability that animals from those colonies would be able to visit a given cell.
This allows, for any colony, predictions to be made for the accessibility of each grid cell in the study region as well as its relative habitat suitability. An advantage of using separate models for accessibility and habitat suitability is that the predictor variables (and the shape of the response curves) can be different in the two models (current and predicted climate change models), allowing more ecologically realistic models .
These accessibility models could be unweighted or weighted by colony size. The weighted models, which considered population sizes of colonies, up-weighted cells in the vicinity of larger colonies and down-weighted cells around smaller colonies thus giving an indication of density. Here, we focussed on unweighted predictions as we cannot know what population sizes will be at the end of the century (2071-2100), the period to which our climate change predictions are forecast, nor how these will be distributed across available breeding locations. However, the predictions based on habitat preference and population density based on current environmental conditions and population sizes are presented in Figure S3 for comparison.

| Mapping
We used our model fits to predict habitat preference for each taxon and to allow comparisons across taxa. However, habitat preference predictions are not absolute estimates of probability and are therefore not directly comparable between species (Beyer et al., 2010).
To allow for comparison between the five taxa, each prediction map was transformed by percentile to give a habitat importance score (Raymond et al., 2015). These final maps present areas of important habitat and are therefore comparable between species, depicting habitat represented on a scale from 0 (not important at all) to 100 (extremely important).

| Redistribution of preferred habitat under climate change projections
To model climate change redistributions, we used two pathways of atmospheric greenhouse gas concentration: RCP4.5 and RCP8.5, which correspond to medium and high radiative forcing, respectively.  (Cavanagh et al., 2017). These projections provide an indication of the redistribution of preferred habitat under the assumption that colonies will not move, and habitat preferences will not change.
In theory, it is possible to use hindcast CMIP5 data and future projections of CMIP5 data and accessibility to predict current preferred habitat and future projections of habitat preference. However, some predictor variables (  (Figure 2). This matched the five most similar cells between the current preferred habitat and current-climate grid, taking into account environmental variables and accessibility conditions.
Projected habitat preference redistributions were then calculated by finding where those same current-climate grid cells could move to, under future projected climate conditions using Euclidean distance.
If the majority of those five cells were from current-climate grid, the future-climate grid cell was classified as "preferred habitat-like" and if not, then "not preferred habitat-like" (Figure 2). This method allows comparison of output from multiple CMIP5 representations, each representing different suites of biophysical variables. Thereafter, the eight redistribution models could be combined into one generalized climate change projection per species and per RCP scenario.
We used a maximum of eight environmental variables per CMIP5 representation (Table S6), depending on the suite of variables available for that representation (i.e. not every CMIP5 representation contained all eight variables). Variables were normalized to range 0-1 prior to further analyses. As the resulting environmental variables showed high correlation, we used a principal components analysis to reduce the number of inter-correlated predictor variables in the data set. The lowest number of principal components required to explain at least 95% of the variance was used.
To assess projected change in habitat area (reduction or increase) per species at a regional scale, we divided the Southern Ocean into the 5 regions using the MEASOshapes package (Brasier et al., 2019) into the Atlantic, Central Indian, East Indian, West Pacific and East Pacific (see Figure S5). We calculated the percentage change between the area of current habitat and the area of predicted habitat for the future (2071-2100) for each of the eight CMIP5 representations to account for the variability between them. We then plotted the degree of change per CMIP5 representation for each penguin species and each region where the species exists to compare the differences in habitat change between scenarios RCP4.5 and RCP8.5.

| Climate change redistributions
Broadly, for all species and across both scenarios (RCP4.5 and RCP8.5), preferred habitat was predicted to move poleward, with habitat being lost in the northern limits of the species' ranges and gained in the southern area of their ranges ( Figure 5). For all rockhopper species, our results predicted a greater average loss of habitat for scenario RCP8.5 than scenario RCP4.5 ( Figure 5).
Overall, there was a projected net loss of preferred habitat for all three rockhopper species (Figure 6). Preferred habitat for eastern rockhoppers showed a mean areal reduction of −1.7% for RCP4.5 or by −4.2% under RCP8.5 ( Figure 6). Northern and southern rockhopper penguins were projected to undergo similar losses of preferred habitat area under RCP4.5 (−6.9% and −6.8%, respectively).

| DISCUSS ION
Here, we identified preferred habitat during the post-moult migra-

| Current habitat preferences
Post-moult habitat preferences for all species largely coincided with the distribution of dominant frontal zones, particularly within open ocean habitat south of the Subtropical Front (STF; Figure 4).
For rockhopper penguins, foraging was generally located within the Subantarctic/Antarctic inter-frontal zone, whereas macaroni/ royal penguins tended to prefer the Polar frontal zone. Latitudinal separation in the use of different inter-frontal zones by penguin species may reflect differences in prey preference but is also seen by the breeding locations of macaroni/royal penguins, which generally breed at higher latitudes ( Figure 1). Resource partitioning mechanisms in areas where species overlap may include vertical spatial partitioning and temporal partitioning such as allochrony (Green et al., 2022; to reduce competition. These findings support previous work documenting strong preferences for frontal zones by marine predators (Bost et al., 2009;Green et al., 2022;.  (Sabine, 2014), which correspond to medium and high radiative forcing, respectively. RCP4.5 and RCP8.5 scenarios depict: Habitat that could remain important (blue), habitat that could be potentially gained (green) and habitat that could be lost (orange).
example, seabird assemblages in the Southern Ocean, including the Eudyptes guild, have previously been associated with specific frontal zones, due to the inferred distribution of prey structured by oceanographic fronts (Bost et al., 2009;Thiebot et al., 2012Thiebot, Lescroël, et al., 2011). The regions that are important for these taxa of Eudyptes penguins occur largely between the south Antarctic Circumpolar Current Front and the STF and are consistent with the distribution of high concentrations of zooplankton (Pinkerton et al., 2020) including euphausiids which are the dominant prey for Eudyptes penguins (Cherel et al., 2007;Rey et al., 2005).

Subantarctic waters (predominantly used by rockhopper species)
are generally dominated by smaller euphausiids such as Euphasia vallentini, whereas larger species such as E. superba (Antarctic krill), E.
frigida and E. triacantha occupy waters south of the Polar Front (Endo et al., 2008), highlighting that these two frontal zones are characterized by different prey size assemblages. The distribution of euphausiids is particularly important for rockhopper species that typically occupy a lower trophic level (Dehnhard et al., 2011). In contrast, macaroni penguins generally consume larger, higher trophic level prey items than rockhoppers including a higher proportion of myctophid mesopelagic fishes (Cherel et al., 2007;. Overall, the importance of frontal zones in controlling biogeography could have considerable implications for prey accessibility for penguins under climate-mediated change. Habitat use by migrating penguins is determined by swimming speed capabilities, currents that assist/hinder movements as well as the profitability of oceanic feeding areas Della Penna, 2016).
Biophysical conditions in the Southern Ocean are changing rapidly (Constable et al., 2014), and whereas cues used by foraging penguins will likely to remain unchanged in the short term (according to generation times of macaroni penguin 8-15 years and rockhopper species 6-10 years), the biophysical features of the ocean and associated locations and availability of their prey may not. Currently, sea surface temperature is changing with the implication that preferred prey species may be pushed farther south or deeper into the water column to follow the cooler waters (Constable et al., 2014;Freer et al., 2019;Veytia et al., 2020).
Results outlined in this study suggest that similar shifts in the distribution of suitable winter foraging habitat could occur for these taxa of Eudyptes penguins as well.

| Potential consequences of habitat redistribution for penguin post-moult foraging ecology
Regionally, interactions of different climate forcings will act on the chemistry and biology of primary and secondary production, with implications for regional food web structure (Sydeman et al., 2015;Trathan et al., 2007). Populations of marine predators from a particular species that breed in different regions may occupy different niches depending on the regionally dominant food web. The blocking cross-validations suggest that this is the case for the eastern and northern rockhopper species due to the low transferability of models between certain populations/species (Torres et al., 2015).
Habitat preferences differ between colonies as there are likely different prey types available, each with their own habitat preferences.
Because of this, marine predator responses to climate-driven change could differ across populations based on variation in prey composition. For example, macaroni penguins consume Antarctic krill as their main prey at South Georgia in the Atlantic (Waluda et al., 2012), while their diet in the Central Indian Ocean is dominated by the crustaceans Euphausia vallentini, Thyssanoesa vicina and Themisto gaudichaudii, followed by myctophid fish and cephalopods (Cherel et al., 2007;Crawford et al., 2003). At Macquarie Island in the East Indian (MEASO-defined region), myctophid fish Krefftichthys anderssoni dominate the diet of royal penguins, followed by Euphausia vallentini (Hull, 1999). These studies show that some species may be able to respond flexibly to changes in the availability of preferred prey. Therefore, while our models predict large-scale redistribu-  (Pütz et al., 2006;Thiebot et al., 2015). Both these areas (pelagic and shelf) coincide with the areas of increased preferred habitat.
Many studies that have investigated how penguin species could respond to a changing climate have focussed on Antarctic breeding, ice-obligate species (however see, Cristofari et al., 2018;Péron et al., 2012). Like Antarctic penguins, changes for Eudyptes penguins relate to trophic mediated changes cascading from climate forcing.
Antarctic penguins have adapted to sea ice conditions (Forcada et al., 2006) and are influenced by the availability of Antarctic krill and the effects of krill harvesting in Antarctic waters. In contrast, Eudyptes penguins are not ice dependent, have a more generalist diet and consume prey species that are generally not commercially valuable. All five taxa of Eudyptes in this study perform post-moult migration round trips of ~10,000 km and cover a large area during the 4-6 months they are at sea. Given their generalist diet and ability to travel large distances, it could be expected that Eudyptes would be better equipped to adjust to environmentally driven changes in the distribution of suitable habitat. However, while regions used by these penguins are vast, key foraging areas used by wide ranging oceanic predators may represent only a small spatial subset of their overall range (Schofield et al., 2010). Areas important for foraging penguins are likely linked to the timing of the annual cycle (Moore et al., 1999), and the spatio-temporal distribution of marine productivity and prey. For example, habitats used by penguins for hyperphagia prior to breeding and moulting are probably disproportionately important and are likely to be relatively close to breeding sites (Thiebot, Authier, et al., 2014;Thiebot, Cherel, et al., 2014).
Changes in the extent and distribution of such habitats progressively away from their breeding colonies could reduce capacity for attaining sufficient condition to survive these energetically costly periods (Dehnhard et al., 2013).
Adjusting to the predicted changes could be also possible for penguins through the establishment of new colonies at sites closer to suitable overwintering habitat. However, data on natal dispersal of juvenile penguins in relation to climate change is currently lacking, and as adults show a high level of philopatry, they have a low emigration rate (Cristofari et al., 2015;Orgeret et al., 2022). Therefore, it is unlikely that changes in preferred habitat in the Southern Ocean would lead to the redistribution of adult birds by emigration, but rather a decrease in populations at sites that become unfavourable and growth of populations at sites that become more favourable.
Where established colonies already exist on islands in the Atlantic and on the Antarctic Peninsula, the increase of preferred habitat to this region could lead to the increase in macaroni penguin populations (as it has for gentoo penguins; Forcada et al., 2006).

| Uncertainties in projected future habitat distributions
Our results highlight sea surface temperature and sea surface height as being the most important predictors of foraging distributions for these species. Together, these variables may reflect broad water masses indicative of preferred prey fields. However, it is possible that these variables simply reflect the broad geographic space indicative of post-moult migration, rather than key characteristics of their foraging habitat. This could have implications when considering how foraging habitat could change in future. In considering the circumpolar average, the Polar Front was reported to have already shifted by 60 km south since 1992 and was expected to continue to do so (Kim & Orsi, 2014;Sokolov & Rintoul, 2009). More recent work has led to the consensus that climate change models do not show a systematic shift in front locations (Chapman et al., 2020;Meijers et al., 2019), but that there is observed poleward warming of the frontal zones (Chapman et al., 2020). The suite of CMIP5 models chosen here all showed strong agreement towards a southward redistribution of penguin habitat. As sea surface temperature may indeed be the most important indicator of penguin foraging, as shown in our study as well (Cristofari et al., 2018;Le Bohec et al., 2008;Morrison et al., 2015;Pütz et al., 2018;Rey et al., 2007), the implications for poleward warming of the frontal zones may be significant enough to have negative effects on widely dispersive penguins. Sea surface temperature affects the locations of foraging for penguins and even short-term poleward shifts of isotherms have resulted in decreases in survival rates and breeding success for penguins (associated with increased energy expenditure when foraging for food; Le Bohec et al., 2008;Trucchi et al., 2019).
Some areas of notable preferred habitat change were found that would be on the extreme edges of species ranges and unlikely to be used by the penguins. These were the notable decrease in preferred habitat for northern rockhopper penguins in the East Indian and southern rockhopper penguins in the West Pacific and increase for macaroni penguins in the West Pacific. Also, at least in the case of the rockhopper species, these areas are relatively small (see Figure S6). Thus, these specific changes, though they seem large in intensity, would likely have only minimal effects to the species.

| Implications and conclusions
While this study does not measure the effects of changing extent of preferred habitat on population numbers, our results, at least in some regions, do seem to coincide with some population trends.  & Moors, 1994;. Notably, while our models predict an increase in preferred habitat for populations from the Falklands, this is not reflected in population trends as long-term declines have taken place from the 1930s to the early 2000s Pütz et al., 2003).
Our models predict that in some cases, shifts south in preferred habitat may occur by the end of the century, forcing some populations to possibly travel farther and spend more energy and time searching and foraging for food. Our habitat predictions only include habitats that are accessible to the penguins currently (according to the calculated distances they may feasibly travel from their colonies). Therefore, the increased habitat in the south of the penguins' ranges is within reach but may be at the upper limit of their energetics to reach.
Ongoing changes in preferred habitat across the Southern Ocean have the potential to result in large scale redistributions of populations like those that took place during the Last Glacial Maximum (c. 19.5-16 kya) after a significant climate warming  and the resultant overall decline in marine productivity (Vianna et al., 2020).
Implications of our results could be that by the end of the century, some populations of the taxa of Eudyptes penguin in this study may decrease to a stable norm of lower numbers and some small existing colonies may start to increase (i.e. the relative importance of colony sites may alter). Our results reiterate previous work showing that responses to environmental change vary across species and between regions. Thus, we emphasize the need for species-specific and regionspecific management and conservation against a backdrop of a rapidly altering marine environment. Finally, our results support those of Gervais et al. (2021), who argued that management needs to be adapted concurrently with changing conditions and should not wait for climate change impact research to justify action.

ACK N OWLED G M ENTS
First and foremost, we thank all the field personnel involved in collecting these data. We also thank the reviewers and editor for their incredibly valuable feedback, which helped improve this