Estimating nest‐level phenology and reproductive success of colonial seabirds using time‐lapse cameras

Collecting spatially extensive data on phenology and reproductive success is important for seabird conservation and management, but can be logistically challenging in remote regions. Autonomous time‐lapse camera systems offer an opportunity to provide such coverage. We describe a method to estimate nest‐level breeding phenology and reproductive success of colonial pygoscelid penguins using photographs from time‐lapse cameras. The method derives from stereotypical patterns of nest attendance, where predominantly two adults are present before and during laying, but switch to one adult during incubation. The switch approximates the date of clutch completion and is estimated by fitting a smoothing spline to daily nest attendance data, identifying candidate dates that switch from two adults to one and selecting the date when the first derivative of the spline is minimized. Clutch initiation and hatch dates are then estimated from the mean, species‐specific interval between laying (pygoscelid penguins typically lay two eggs) and the duration of the incubation period. We estimated these intervals for each species from historical field data. The phenology is adjusted when photographs indicate egg or chick presence prior to their estimated lay or hatch dates. The number of chicks alive in each study nest on its crèche date determines reproductive success estimates. The method was validated with concurrent direct observations for each species and then applied to a camera network in the Antarctic Peninsula region to demonstrate its utility. Mean egg laying and incubation intervals from direct observations were similar within species across sites. In the validation study, the mean clutch initiation, hatch and crèche dates were generally equivalent between photographs and direct observations. Estimates of reproductive success were identical. Applying the method to a time‐lapse network suggested relatively high reproductive success for all species across the region and corroborated general understanding of latitudinal trends and species‐level plasticity in phenology. The method accurately estimated phenology and reproductive success relative to direct observations and appears well‐suited to operationalize regional time‐lapse camera networks. The estimation method should be applicable for other seabirds with stereotypical nest attendance patterns from which breeding phenology could be estimated.


| INTRODUC TI ON
Data on reproduction, including breeding phenology and nest success, are important for management and conservation of seabirds globally (Cairns, 1987;Constable, de la Mare, Agnew, Everson, & Miller, 2000;Cury et al., 2011). In particular, breeding phenology and reproductive success data from seabirds are thought to indicate general conditions of marine ecosystems (Cairns, 1987) and represent important components for the development of ecosystem-based fisheries management (Einoder, 2009). However, monitoring to collect such data can be time intensive and requires experienced field personnel, often in remote sites for extended periods of time. The commitment of personnel to the field may also necessitate spatially restricted data collection (relative to species distribution). In particular, long-term ecological studies at single sites can provide high resolution data, but population-level inference requires an assumption that such data represent regional trends. This is not always the case (e.g., Lynch, Naveen, Trathan, & Fagan, 2012). Furthermore, direct observations of nests may bias estimates of reproductive phenology or success via a variety of mechanisms (e.g. observer disturbance, predator facilitation, nest abandonment) that can ultimately result in nest failure (Carney & Sydeman, 1999). As an alternative, autonomous camera networks may provide a solution that can expand spatial coverage of seabird monitoring in a cost-effective, non-invasive way (e.g. Newbery & Southwell, 2009). Widespread application of such systems would benefit from simple methods to standardize analysis of data derived from photographic images. We report a novel method to estimate seabird breeding phenology and reproductive success using colonial pygoscelid penguins as a model, with photographs collected from time-lapse cameras. We apply the method to a collaborative, multi-national camera network that was deployed in the Antarctic Peninsula region in the austral summer of 2015/16 to monitor penguin colonies.
Remote photography, defined as "photography or videography of wild animals in the absence of the researcher" (Cutler & Swann, 1999), is commonly used for research and monitoring of seabirds around the world (Cutler & Swann, 1999), particularly for studying nest predation (e.g. Collins, Green, Dodd, Shaw, & Halsey, 2014;Davies, Dilley, Bond, Cuthbert, & Ryan, 2015), nesting activity (e.g. Weller & Derksen, 1977) patterns of attendance (e.g. Black, Collen, Johnston, & Hart, 2016;Huffeldt & Merkel, 2013;Lynch, Alderman, & Hobday, 2015;Southwell & Emmerson, 2015;Southwell et al., 2013) and to estimate reproductive success (e.g. Merkel, Johansen, & Kristensen, 2016). Another potential application of these systems is to use time-lapse cameras to estimate the breeding phenology (i.e. the timing of reproductive events, including dates of clutch initiation, hatch and crèche) and reproductive success (i.e. numbers of chicks raised to independence per nest). Such data are useful for examining factors that impact seabird populations, including climate change (e.g. Visser & Both, 2005) and fishing (e.g. Agnew 1997;Constable et al., 2000;Cury et al., 2011). For many species, the timing of phenological events (e.g. laying or hatching) can vary inter-annually and spatially depending on local environmental conditions, but the duration of intervals between specific phenological events (e.g. duration of incubation) tend to be more fixed. Thus, estimating annual breeding phenology minimally requires identifying a reliably observed event that can be placed into a known timeline, allowing back or forward calculation of the dates of other unobserved events. Estimating breeding phenology and reproductive success from time-lapse photography among colonially nesting pygoscelid penguins may be particularly ideal, given stereotypical patterns of adult attendance at their nest during the breeding season, relatively fixed periods of time between events in the breeding cycle and fidelity of chicks to their nest from hatch until crèche (defined here as the day when the chick is first left unattended by a parent). Such characteristics provide observable indicators of major events during the breeding season from which breeding phenology and reproductive success may be estimated, even if nest contents or particular breeding events cannot be observed directly or regularly in photographs. Southwell and Emmerson (2015) demonstrated that direct observations. Estimates of reproductive success were identical. Applying the method to a time-lapse network suggested relatively high reproductive success for all species across the region and corroborated general understanding of latitudinal trends and species-level plasticity in phenology.
4. The method accurately estimated phenology and reproductive success relative to direct observations and appears well-suited to operationalize regional time-lapse camera networks. The estimation method should be applicable for other seabirds with stereotypical nest attendance patterns from which breeding phenology could be estimated.

K E Y W O R D S
Antarctica, camera, monitoring, penguin, phenology, reproductive success, seabird, time-lapse peak attendance of adults at the colony level was synchronized at the start of laying in Adélie penguins Pygoscelis adeliae. Here, we extend this idea to the nest level and develop a simple method to reconstruct breeding phenology from photographic observations of nest attendance and opportunistic verification of nest contents.
In photographs, the number of parents attending the nest and the presence/absence of large chicks are reliably observed (Supporting Information Figures S1-S3). Direct observations of specific phenological events (i.e. lay and hatch) are possible less frequently, because protective postures by adults generally preclude a clear view of nest contents in photographs. However, among pygoscelid penguins, clutch completion is typically marked by a shift in adult attendance at the nest from predominantly two birds to predominantly one (e.g. Trivelpiece & Trivelpiece, 1990;Williams, 1995). This shift is readily observable because mates alternate incubation duties to forage at sea. If the date of this shift in adult attendance is estimable, then the interval between laying (pygoscelid penguins typically lay a maximum of 2 eggs per nest) can be used to back-calculate lay dates and the duration of the incubation period can be used to forecast hatch dates. Breeding phenology and success of the nest can then be completed with observations of the crèche date and the number of chicks alive on that day respectively.
To advance the use of time-lapse cameras to provide standardized data on phenology and reproductive success, we report on: (1) a method for estimating breeding phenology from photographic records of adult attendance and nest contents at focal nests; (2) mean durations of the laying, incubation, and brood/guard periods that are necessary to parameterize the estimation procedure for Adélie, chinstrap P. antarctica and gentoo P. papua penguins from several monitoring sites around Antarctica; (3) validating the estimation method with direct observations collected concurrently for each species; (4) applying the estimation method to a remote camera network newly deployed in the Antarctic Peninsula region; and (5) a sensitivity analysis to identify the minimum number of daily timelapse images necessary for confidence in the estimated phenology.

| Camera deployment
We used autonomous time-lapse cameras (Reconyx Hyperfire HC500 or PC800) with an expected operational endurance of greater than 1 year when deployed with 12-AA lithium metal batteries. This endurance was essential, as visits to some sites are only possible for short periods once per year. Cameras were deployed to capture a minimum of 6-12 photographs per day, taken at 30 min or 60 min intervals, between local daylight hours of 09.00 and 15.00.
The cameras were positioned 1.5-2 m above-ground level on tripods or partially buried metal poles. In general, focal nests were 2-10 m from the camera and photographed at an oblique angle (between 8° and 45°) to facilitate viewing of nest contents. For each camera, experience suggests that up to 20 nests can be reliably monitored for the duration of the breeding season depending on nest density and topography. The cameras were deployed at several sites along the South Shetland Islands and Antarctic Peninsula (Figure 1). Data for the validation study (see below) were collected at Cape Shirreff, Point Thomas and Lion's Rump (Table 1). An example camera deployment is shown in Figure S4.

| Photo classification
Photographs were classified manually by teams from each site using the following protocol. Nests were selected for daily classification by identifying those which contained at least two adult birds prior to laying. From the daily set of available photographs, the maximum adult attendance at each study nest (nest attendance was defined simply as the number of adults associated with a given nest and this can be visualized in Figures S1 and S2) was recorded beginning on the date when two adults were observed attending an empty nest bowl. Daily classification of nest attendance and nest contents proceeded until nest failure or crèche was confirmed.
Nest contents (the number of eggs and chicks) were identified and counted only when clear evidence of their presence or absence was visible in a photograph. On days when the nest was not visible due to nest obscurement by other birds in the foreground, iced lenses or poor visibility due to storms, fog or heavy precipitation, nest attendance was recorded as unknown. On average for each nest, all-day obscurement occurred 2% of the time across the camera network. Photographic evidence confirming a lay or hatch event was also recorded. If one photograph exhibited no eggs or chicks in the nest, but a subsequent photograph within 24 hr revealed an F I G U R E 1 Map of the camera network sites for Adélie (black), gentoo (red) and chinstrap (blue). Inset shows the study location (red shaded polygon) relative to Antarctica TA B L E 1 Data summaries from time-lapse studies during the 2016/17 field season by species and locality from the remote camera network in the Antarctic Peninsula region. Due to logistic necessity, data retrieval from the cameras at Paradise Bay occurred prior to crèche. Validation sites for each species are noted with an italicized locality egg or chick, the lay or hatch date, respectively, was registered as the day on which the egg or chick was observed. Similarly, the presence of crushed or partial egg shells on the nest was considered evidence of hatch, as those shells are typically ejected from the nest bowl following hatch and quickly lost to predators, winds or trampling. Crèche dates were recorded on the date when the clear association between a parent and its chicks at the nest was not distinguishable or when chicks were clearly unattended in their nest. Note that identifying the crèche date, both on the ground and in photographs, is nonetheless subjective because, without an identifying mark, movement of birds and temporary associations of chicks with other birds in the colony at this time hinder definitively tracking parent-offspring associations. We briefly discuss the time investment for manual classification of photographs later.

| Estimation of clutch initiation and hatch dates
Nest-level clutch initiation dates (CID, the date when the first egg was laid) and hatch dates were estimated from the photographic attendance and nest content data with a four-step process. Our approach assumes that nest attendance during daylight hours exhibits a switch from predominantly two birds to predominantly one bird around the time of clutch completion (Trivelpiece & Trivelpiece, 1990). The date of this shift in attendance was estimated by first fitting a smoothing spline (Chambers & Hastie, 1992), implemented with the smooth.spline function in r (R Core Team, 2016) with 10 df, to the attendance data and taking the first derivative of the fitted smooth. Next, the attendance data were differenced (lag of 1) to identify candidate dates when the observed nest attendance shifted from two to one. The switch date was selected from the candidate dates where the first derivative (slope) of the smooth was minimized.
The estimation procedure is illustrated in Supporting Information Figure S5. The CID was then back-calculated from the shift date based on a species-specific mean interval between the first and second lay dates (see section below). Direct observations of nest attendance during the laying period suggest that the switch date generally occurs at the time of clutch completion by Adélie penguins, up to 1 day prior to clutch completion by chinstrap penguins, and 1 or 2 days before clutch completion by gentoo penguins (Trivelpiece & Trivelpiece, 1990). This apparent switch prior to clutch completion owes to daytime foraging of one of the mates, with subsequent returns for either clutch completion or incubation relief. We therefore adjusted the back-calculation of CID from the switch date by 0 days for Adélie, and 1 day for chinstrap and gentoo penguins to account for these stereotypical attendance patterns during the laying period (Trivelpiece & Trivelpiece, 1990). The hatch date for the first chick was then projected from the estimated CID based on a species-specific mean incubation period based on historical direct observations (see section below).
The CID and hatch dates estimated from attendance data were checked against the nest content observations and adjusted if necessary. First, if an exact lay date was observed in the photographs, we replaced the estimated CID with the observed CID.
Second, if an egg was observed in the nest prior to the estimated CID, we back-calculated a new CID from the first egg observation date. This calculation was based on the validation data (see below) which suggested that the first observation of an egg in a photograph occurred, provided the egg was observed within 1 week of the true CID, 2 ± 1.8 (SD) days after true clutch completion. Hatch dates were recalculated for any corrected CID. Finally, we replaced the estimated hatch date with the observed hatch date if the hatch date was considered known. The code for this estimation procedure was developed in r v 3.2.2 (R Core Team, 2016) and is available in Supporting Information Appendix S1.

| Sensitivity analysis
The sensitivity of the estimated switch date to photograph frequency (number of photographs per day) and interval (time elapsed

| RE SULTS
Historical data on laying and incubation intervals were similar across species and sites. The mean laying interval was approxi-  F I G U R E 3 Boxplot of nest-level differences (in days) between photograph-estimated and directly observed dates for clutch initiation (CID), hatch and crèche for Adélie (black circles), gentoo (red triangles) and chinstrap penguins (blue squares) combined. The whiskers represent the full range of observed differences. The mean difference in Adélie crèche date estimates is indicated with a dashed black line. The dotted line marks a difference of 0 days for reference regression models indicated that latitude explained a large proportion of the variation in CID for Adélie (R 2 = .99, F 1,3 = 602, p < .01) and gentoo penguins (R 2 = .74, F 1,5 = 14.3, p = .01), while chinstrap CID was not explained by colony latitude (R 2 = .38, F 1,3 = 1.82, p = .27), noting that the latitudinal range of chinstrap observations was smaller than for the other species (Figure 4a). The timing of phenological events across the network exhibited species-specific windows (Figure 4b), with gentoo penguins exhibiting the widest range of dates for CID, hatch and crèche, while each breeding phase for chinstrap penguins showed little spatial variation.
The sensitivity analysis suggested that, for both 30-and 60-min intervals, the switch date is well estimated by ≥4 consecutive photographs per day ( Figure 5).

| D ISCUSS I ON
The

| Application to a camera network
Applying the method to a recently deployed network of timelapse cameras demonstrated its utility across monitoring sites and research teams, and, importantly, provided results consistent with known phenological variation due to colony latitude and plasticity among the pygoscelid penguins breeding in the Antarctic Peninsula region (Black, 2016;Hinke, Polito, Reiss, Trivelpiece, & Trivelpiece, 2012;Lynch, Fagan, Naveen, Trivelpiece, & Trivelpiece, 2009). In particular, the relatively high degree of plasticity in gentoo penguins relative to Adélie penguins has been shown for inter-annual differences in phenology (Hinke et al., 2012;Juáres et al., 2013). This work extends that result to suggest intra-annual plasticity is also higher in gentoo penguins than in Adélie penguins. Similarly, while inter-annual variation in breeding phenology of chinstraps can be high (e.g. Black, 2016), the narrow window of time for each phenological event exhibited by chinstrap penguins across the camera network in 2016/17 was an unexpected, novel result. Chinstrap penguins are highly migratory (e.g. Hinke et al., 2017) and, like other migratory species, their arrival to the colony and subsequent breeding phenology may not be as strongly coupled or sensitive to local breeding conditions (Both et al., 2010). However, the interactions between migration triggers and local breeding conditions that might allow inter-annual variation in phenology (Black, 2016) but little intra-annual spatial variation as observed here remain unclear. Further monitoring will be worthwhile to assess the generality of this novel result among chinstrap penguins. With respect to the method developed here, it is evident that integration of data collection and analysis methods can provide novel insights on spatial scales beyond focal colony monitoring.

| Advantages and disadvantages
The phenological estimation method based on time-lapse data has several advantages relative to traditional direct observations. One major advantage is the capacity to include monitoring at colonies that researchers are unable to regularly visit, as a camera can run for at least one full year without maintenance. Solar-powered options, such as those described by Newbery and Southwell (2009)  contents to be reliably identified. The cameras we used provided dpi, ca. 500 kb) that was adequate for data needs, but higher resolution or larger format photographs (e.g. Lynch et al., 2015;Southwell & Emmerson, 2015) could also be used. Second, as noted above, roughly 80% of nest-level phenologies were estimable from adult attendance data only. A primary data requirement for adult attendance may improve efficiency of data collection from images to support the estimation procedure. Nonetheless, we urge consistent identification of nest contests, as these are necessary to relax fixedinterval constraints that might mask inter-annual or spatial variation in breeding chronologies (e.g. Black, 2016). Finally, the restricted nature of the attendance data (e.g. 0,1, or 2 adults) facilitates the use of a simple statistical method to estimate breeding phenologies while accounting for uncertainty and variability in nest attendance patterns around the time of clutch completion (e.g. Figure S5).
Finally, since images for this analysis are taken during daylight hours, timing and frequency of the photographs can be optimized to achieve results without excessive picture accumulation. Prior knowledge of attendance patterns, both seasonal (Southwell et al., 2013) and diurnal (Merkel et al., 2016), would aid the design of appropriate sampling protocols for other colonial species. Necessary photographs also could be extracted from higher frequency image collection protocols if other breeding season parameters were prioritized. Foraging trip durations, incubation shifts or diurnal attendance patterns could all be estimated based on adult attendance of the nest (Huffeldt & Merkel, 2013;Lynch et al., 2015), and the phenology could be estimated from a subset of higher frequency photographs collected during the day. Thus, multiple datasets could be collected from the same images, further enhancing the efficiency of remote camera networks to provide spatially extensive monitoring data. Finally, the phenology estimation method requires an image at the beginning of the breeding season that clearly identifies two adults and an empty nest bowl. This constraint ensures standardization of data collection. However, confirmation of this condition can take time due to large numbers of individuals in a colony, ongoing nest construction and sometimes poor conditions within the colony, such as excess snow or wet guano, that hinder identification of a nest bowl. However, once the nest location and association is determined, classification proceeds quickly.

| CON CLUS ION
The use of autonomous data collection systems is rapidly growing in the field of wildlife biology and ecology. As the use of autonomous systems increases, standardized methods for data collection and analysis will help ensure compatible results and foster collaborations. The estimation method described here appears well-suited to operationalize regional applications of time-lapse cameras to estimate phenology and reproductive success of wild pygoscelid penguins, a focus of ecosystem monitoring efforts in the Southern Ocean (Agnew, 1997). However, while this method was developed for and tested on pygoscelid penguins, it should apply generally to other large-bodied (e.g. >1,000 g) colonial seabirds that mate at the nest site and alternate incubation duties after clutch completion.
Examples include, but are not limited to, ground-or cliff-nesting seabirds such as albatross (Diomedeidae), giant petrels Macronectes spp. and other fulmarine petrels (Procellariidae), boobies (Sulidae), many gull (Laridae) and cormorant (Phalacrocoracidae) species, murres Uria spp., as well as other penguin species including rock hopper Eudyptes chrysocome and E. moseleyi and macaroni penguins E. chrysolophus. Careful consideration must be given to selecting suitable aggregations of nesting seabirds for automated observation systems because terrain, vegetation and nest density may limit the efficacy of camera systems to provide such data.
However, if attendance data are available, adapting the estimation method to other species or locations would simply require data on species-specific phenological intervals and stereotypical nest attendance patterns for the species and region of interest.

ACK N OWLED G EM ENTS
We