Large‐scale assessment of intra‐ and inter‐annual breeding success using a remote camera network

Abstract Changes in the physical environment along the Antarctic Peninsula have been among the most rapid anywhere on the planet. In concert with environmental change, the potential for direct human disturbance resulting from tourism, scientific programs, and commercial fisheries continues to rise in the region. While seabirds, such as the gentoo penguin Pygoscelis papua, are commonly used to assess the impact of these disturbances on natural systems, research efforts are often hampered by limited spatial coverage and lack of temporal resolution. Using a large‐scale remote time‐lapse camera network and a modeling framework adapted from capture‐recapture studies, we assess drivers of intra‐ and inter‐annual dynamics in gentoo penguin breeding success across nearly the entire species’ range in the Atlantic sector of the Southern Ocean. We quantify the precise timing of egg/chick mortality within each season and examine the role of precipitation events, tourism visitation, and fishing activity for Antarctic krill Euphausia superba (a principal prey resource in the Antarctic) in these processes. We find that nest failure rates are higher in the egg than the chick stage and that neither krill fishing nor tourism visitation had a strong effect on gentoo penguin breeding success. While precipitation events had, on average, little effect on nest mortality, results suggest that extreme weather events can precipitate sharp increases in nest failure. This study highlights the importance of continuous ecosystem monitoring, facilitated here by remote time‐lapse cameras, in understanding ecological responses to environmental stressors, particularly with regard to the timing of events such as extreme weather.

Each image captured over the course of the season was uploaded to the Zooniverse platform ( Fig. S1-2). We classified the location of penguin chicks in each image obtained from the remote time-lapse camera network from the sighting of the first chick at a site to the crèche stage (the period during which penguin chicks begin to spend less time at the nest and form large chick aggregations). The zones were used as guidelines for users, as chicks from one nest may slightly stray into an adjoining zone. We did not mark any chick where there was any ambiguity regarding which nest it belonged to, as false positives (in the form of incorrect nest membership for a given chick) would bias estimates in this modeling framework. Ambiguous nest membership was rarely an issue in scoring and all classifications were reviewed for accuracy to avoid potential false positives. Chick body position and proximity to nest were used to distinguish live from dead chicks (only live chicks were scored), though there was rarely any ambiguity. This method effectively provides a time series of the number of observed chicks for each nest/site/year ( Fig. 1b in main text).

Data Availability
Table S1-1 : Breeding success estimates from this study and estimates compiled from two other studies in the literature (providing an additional 11 site/years of information). Estimates from other studies are provided for reference, though these data were not included in any analyses. Note that measures of uncertainty for estimates from this study (given as posterior standard deviations) are not provided for estimates from other studies in the literature.

Site Name
Year Latitude Longitude Chicks / Pair Uncertainty (sd) Source

Posterior predictive check -capture-recapture model
Posterior predictive checks were used to assess the degree to which the model could produce data that resemble the observed data. The test statistic chosen was the total number of observed chicks for each site/year. Bayesian p-values represent the proportion of generated datasets with a test statistic (in this case, the total number of chicks 'observed') that is greater than the test statistic derived from the actual data. Bayesian p-values close to zero or one indicate that the model generates data different than those used to fit the model. Results did not indicate any model misfit (as evidenced by Bayesian p-values near 0.5).

Precipitation coding
Figure S1-4 : Example images used for precipitation coding. The top row represents rain events, while the bottom row represents snow events. For each precipitation type, the magnitude of the events increases from left to right (R1/S1 left, R2/S2 middle, R3/S3 right).

Effect of precipitation events on the timing of nest failure
While smaller precipitation events were of little consequence for chick mortality (Fig S1-5), large precipitation events appear to impact breeding success in a non-linear fashion (Fig. 5, Video S1). Extreme weather events can result in increased egg failure or chick mortality.
For the plots below, mean detection probability at each site in every year at every time step p t,j,k was calculated as a derived quantity by averaging across the detection probabilities p t,i,j,k for all nests i, i,j,k where N is the number of nests for a given site and year.