Projected asymmetric response of Adélie penguins to Antarctic climate change

The contribution of climate change to shifts in a species’ geographic distribution is a critical and often unresolved ecological question. Climate change in Antarctica is asymmetric, with cooling in parts of the continent and warming along the West Antarctic Peninsula (WAP). The Adélie penguin (Pygoscelis adeliae) is a circumpolar meso-predator exposed to the full range of Antarctic climate and is undergoing dramatic population shifts coincident with climate change. We used true presence-absence data on Adélie penguin breeding colonies to estimate past and future changes in habitat suitability during the chick-rearing period based on historic satellite observations and future climate model projections. During the contemporary period, declining Adélie penguin populations experienced more years with warm sea surface temperature compared to populations that are increasing. Based on this relationship, we project that one-third of current Adélie penguin colonies, representing ~20% of their current population, may be in decline by 2060. However, climate model projections suggest refugia may exist in continental Antarctica beyond 2099, buffering species-wide declines. Climate change impacts on penguins in the Antarctic will likely be highly site specific based on regional climate trends, and a southward contraction in the range of Adélie penguins is likely over the next century.

Projected asymmetric response of Adélie penguins to Antarctic climate change Megan A. Cimino, Heather J. Lynch, Vincent S. Saba, Matthew J. Oliver

Supplemental Figures
Supplemental Figure 1. The distribution of sea surface temperature (SST) and sea ice concentration at increasing or stable and decreasing Adélie penguin colonies from 1981-2010.  Novel Climate due to: Bare rock Adelie colony (n=136)  Novel Climate due to: Year(((( Increasing Year(((( Novel Climate due to:  Novel Climate due to:     Tables   Supplemental Table 1. Climate models and groups from the Intergovernmental Panel on Climate Change (IPCC) assessment report (AR5) and two additional NOAA GFDL models that we used in our analyses. For both modeling approaches, we used a cross-validation resampling procedure with four replicate runs that partitioned 75% of the penguin colonies into the fitting fold and 25% of the colonies into the left out fold 1 . This allowed for assessment of predictive performance on the held-out folds using the area under the receiver operating characteristics curve (AUC). The AUC is an indicator of the accuracy of the models, where 1 represents a model with perfect performance and 0.5 indicates a model that is no better than random 10 . In MaxEnt, jackknife tests were used to quantify which environmental predictors are contributing the most to fitting the model. For GAMs, we estimated the importance of each predictor variable as described by 11 . In all models, the species prevalence was set to 0.147, which is the true prevalence of Adélie penguins in the Southern Ocean on our polar stereographic grid. We compared MaxEnt and GAM predictions (Supplemental Fig. 25), fitted response functions (Supplemental Figs. 12, 26), variable importance and AUC (Supplemental Table 2).

Matching novel climate and chick-rearing habitat suitability to penguin colony locations
From 1981-2010, we matched predicted trends in CRHS from MaxEnt and GAMs to colony locations and compared trends in CRHS in different Antarctic sectors. We also determined the number of years with novel climate for each coastline pixel, which was also matched to documented population trends. If novel climate occurred at a colony with a population status, we determined the main cause for that novel climate: warm SST, cool SST or high SIC (low SIC was not a category because the lowest SIC (zero sea ice) was documented).
We used a nonparametric Kruskal-Wallis test to determine if there were significant differences between documented penguin population statuses and the number of years with novel climate.
We also used a multiple comparison test after Kruskal-Wallis to determine if the number of years with novel climate differed between population groups. For future climate projections, we matched novel climate to current colony locations to understand how conditions at those colonies could change in the future.

Supplemental Results
To verify and understand the sensitivity of species distribution model predictions (Fig. 2), we tested different combinations of PA datasets (all = full dataset, Cont = only continental locations; Fig. 1) and used two species distribution model approaches (MaxEnt and GAMs) (Supplemental Fig. 25). Mean CRHS from respective GAM and MaxEnt models were highly correlated (Pearson correlations, r > 0.95, p < 0.05). We compared trends in CRHS from respective GAM and MaxEnt models and found ALL PA and Cont P/All A were highly correlated (r > 0.90, p < 0.05), Cont P/A were significantly correlated (r = 0.71, p < 0.05), and All P/Cont A had a lower, yet significant correlation (r = 0.39, p < 0.05). A noticeable difference between MaxEnt and GAMs appears in the Cont P/A example in which MaxEnt produces results more similar to All P/A compared to the respective GAM. In general, the high correlations between MaxEnt and GAM predictions agree with other studies demonstrating the high correlation between presence-only MaxEnt and PA GAM results 12 but we also show that model results are more similar when PA data is complete. All models performed well (area under the curve (AUC) > 0.85, Supplemental Table 2) and predicted trends in suitability were higher at locations with available bare rock, which is necessary for nesting (Supplemental Fig. 27, 28).
The trends in CRHS by Antarctic sector at present, absent, and no bare rock locations highlight how model predictions deviate based on given PA data (Supplemental Fig. 28). Predictions can be further explained by response curves (Supplemental Fig. 12, 26) and variable importance (Supplemental Table 2).
The WAP is a warmer environment compared to the continent (Fig. 1), and thus, represents a different environmental niche that Adélie penguins occupy. Models that did not include WAP absence or pseudo-absence data over-predicted mean and trends in CRHS in the WAP (ex. Cont P/A and more so for All P/Cont A, Supplemental Fig. 25). The response curves show a change in the right hand tail of the SST distribution when WAP absence data was excluded (Supplemental Fig. 12, 26). SST and SIC observations that were outside the range of the model training data, caused the models to extrapolate into novel climate. For All P/Cont A, excluding absence data along the warmer WAP resulted in higher suitability predictions for all WAP locations because the model has information that the penguin colonies are present within this SST and SIC range. Furthermore, losing absence data across part of the range (All P/Cont A) appears more harmful than losing presence data across part of the range (Cont P/All A). This can be seen in the mean CRHS and trends in CRHS in which CRHS did not substantially change