Global analysis of emperor penguin populations

Description

Notes

Methods

During the 2018 Antarctic field season, under permit #2019-006 granted by the National Science Foundation, our US-based team conducted aerial photography at emperor penguin colonies in the Ross Sea to add to robust validation of imagery. Our efforts included one flight via fixed wing aircraft over colonies distant from McMurdo Station and five flights via helicopter to a single colony (Cape Crozier) near the station. The five flights to Cape Crozier, 24 October to 15 November, were used to better understand population fluctuation through a single season. Our fixed wing survey took place on 31 October 2018, flying in the vicinity of Beaufort Island (ASPA 105), Franklin Island, Cape Washington (ASPA 173), Coulman Island, and Cape Roget. At each location (both by fixed wing and helicopter), we circled the colony 1-4 times, maintaining a minimum of 500 m horizontal distance from the periphery of the colony and a minimum altitude of 500 m. No behavioral disturbance to birds (e.g., rapid movement or dispersion of adults or chicks) were noted during any flight. Oblique photographs (with a Canon EOS Mark 7D II with Tamron 400m zoom lens) were taken through the window of the Basler aircraft, and in the case of our AStar helicopter surveys with the window down, in continuous shooting mode to ensure effective ability to stitch photos together for counting.

                We then downloaded and stitched the multiple photos per colony with Adobe photo-stitching software to create a single image for manual counting. We loaded images of colonies into the free software ImageJ, which allowed us to document and assess precision of counts among observers. Our field team (four people) counted the largest colony in the world, Coulman Island, to gain an understanding of among-counter precision. After determining a coefficient of variation of ~2.5% (small variation in counts among observers), we determined that each team member would be assigned to count one of the remaining colonies each, to speed the process and to arrive at a population count of adult emperor penguins at each of six Ross Sea colonies during spring 2018. These data were immediately entered into MAPPPD repository. We used these counts as validation for our observation model (see population modeling below). 

Satellite imagery

                 To gather images for analysis, we first reviewed discover.digitalglobe.com (Maxar Technologies) to determine image availability per emperor penguin location per year, and to determine utility/quality of images for analysis. Images had been requested for acquisition via the National Geospatial Intelligence Agency (NGA), when possible, once per month during cloud-free days in austral spring. We avoided images with excessive clouds, and those that were too dark, too bright, or otherwise low-quality. We created a list comprising the unique identifier for each image (called the "catalog ID") and then requested images be processed (via Polar Geospatial Center [PGC]), specifically pan-sharpened (i.e., increasing the spatial resolution of the multispectral image by merging it with its higher-resolution, panchromatic pair) and projected to Antarctic Polar Stereographic (ESPG code 3031). We then followed semi-automated methods already established: briefly, we loaded VHR imagery into ArcGIS 10.8 (ESRI), identified the location of emperor penguins on the image, and then clipped images to the extent of the colony. We conducted a supervised classification by manually training the program with shapefile points representing pixels of guano, snow, and penguin. We conducted a maximum likelihood classification based on these classes to arrive at a classified raster image identifying pixels that are likely penguin pixels. Our final step was to convert the raster to a polygon shapefile and to calculate the area (m²) of penguin pixels. The area of "penguin pixels" per colony per year served as the response variable and input to the population modeling (below). We conducted this process for all 50 colonies in all years for which imagery existed during 2009-2018.

Model overview

We developed a Bayesian state-space model to accommodate several key features of emperor penguin population dynamics and the data collection (i.e., observation) process. We gathered all available adult count data (obtained from remote cameras, ground, and aerial surveys) from MAPPPD for colonies that ranged in size and in proximity to research stations, and which were situated in different regions of the Antarctic.

The population processes we sought to model were:

1) Colony-level trends and annual fluctuations can differ, even among nearby colonies;

2) Individual colonies can temporarily "vanish" for a breeding season and re-appear in future years (somewhat depending on fast ice conditions);

3) Daily abundance of adult penguins in a colony can vary substantially throughout the spring (August – December) survey period, caused by breeding synchrony, temporal variation in adult foraging trips, the presence/absence of non-breeding adults, emigration from the colony, breeding failure, and changes in parenting behavior (crèching). As noted, unknown to us was the prevalence of these factors before images were acquired, thus subsequently affecting what we measured as "colony size".

The data collection (i.e., observation) processes we sought to model were:

4) Counts of adults from aerial surveys are an imperfect observation of the seasonal population index (i.e., due to counting errors during surveys and intraseasonal variation in daily abundance of adults at colonies; point 3 above);

5) Satellite observations of the "area of ground occupied by penguins" are imprecise and potentially biased estimates of the true count during the survey, and by extension, of the seasonal expected count;

6) The expected number of birds counted at a colony (either through aerial surveys or satellite images) potentially changes over the survey period, owing to chick mortality and subsequent emigration by attendant adults.

Model fitting

All data were analyzed using R version 4.2 (R Core Team 2021), with posterior samples generated using Markov chain Monte Carlo methods implemented using JAGS version 4.3. After a burn-in period of 50,000 iterations, we stored every 50th iteration until we accumulated 10,000 posterior samples from each of three Markov chains. The model unambiguously converged; the Gelman–Rubin convergence statistic was less than 1.1 for all hyperparameters, colony- and year-level effects, regression coefficients, and latent states. We confirmed that the effective sample size for each parameter was greater than 2,000 and also confirmed the ability of the model to generate data that are consistent with the observed data, using posterior predictive checks. We confirmed the ability of our model to generate unbiased and identifiable trend estimates using simulations. We report the medians and 95% equal-tailed credible intervals of all modeled quantities unless otherwise noted. Code and data to replicate our analysis is available at https://github.com/davidiles/EMPE_Global.

Goodness-of-fit and model diagnostics

Posterior predictive checks confirmed that the fitted model could generate data with reasonable properties and no obvious systematic discrepancies with the observed data (i.e., data simulated from the fitted model "looks like" the observed data). Of our simulated datasets, 31% had aerial observations with lower RMSE than the observed data (i.e., Bayesian p-value = 0.31) and 29% of simulated datasets had satellite observations with lower RMSE than observed data (i.e., Bayesian p-value = 0.29; Figure S1). Bayesian p-values close to 0.5 indicate a reasonable fit and occur when the fitted statistical model is equivalent to the "true" model that generated the data. Visual inspection of observed versus fitted values (Figure S2) also indicated the model was a good fit to the data.

Simulations to confirm parameter identifiability

We conducted a series of simulations (n = 300) to confirm that the statistical model could generate identifiable and unbiased estimates of global population trend and change, given realistic data availability, observation error, and survey imbalance among colonies. In each simulation, we generated a time series of "true" abundance at each of the 50 colonies (also resulting in a simulated global trend), and then simulated aerial and satellite observations at each colony, including realistic data imbalance, as well as aerial and satellite observation error (based on values estimated from the analysis of empirical data; Table S3). We then used these simulated observations as "data," re-fit our statistical model to those simulated data, and evaluated whether we could recover unbiased estimates of global trend with appropriate credible interval coverage (Figures 1 and 2). Simulations indicated that the model could reliably recover estimates of global trend (median bias = 0.3%; Figure 2) and change between 2009 and 2018 (median bias = 3.5%; Figure 2), with appropriate 95% credible interval coverage. These simulations suggest that the extreme data imbalance in our data, coupled with our choice of priors, does not induce severe bias into estimates of population change.

Sea ice correlations

To investigate a possible relationship between regional trends in fast ice or pack ice, we gathered published data on fast ice trends and pack ice trends within discrete regions of Antarctica that differ in their patterns of ice formation and sea ice co-variability. We then assigned each emperor penguin colony to these sea ice regions. Within each region, we used samples from the Bayesian joint posterior to sum colony indices and thereby calculate estimates of regional indices and population change. Finally, we estimated the Spearman rank correlation between regional population and regional trends for fast ice (n = 8), and pack ice (n = 5).