Arctic-nesting barnacle geese are divided into three flyway populations, with geese breeding in Eastern Greenland, Svalbard and along the Barents Sea coast; wintering in Ireland, the UK and the Netherlands/Germany, respectively (Madsen et al. 1999, Fig. 2a). Between 2000 and 2016, barnacle geese were studied in three breeding colonies, of which two are located in the high Arctic (Svalbard) and one in the low Arctic (at the Russian coast of the Barents Sea). (1) On the islet Storholmen in Kongsfjorden (KF), Svalbard (78°55′N, 12°12′E, Fig. 2b); (2) on the islet Diabasøya and adjacent tundra at Nordenskiöldkysten (NSK), Svalbard (77°46′N, 13°42′E, Fig. 2b) and (3) surrounding the abandoned village of Tobseda at the Kolokolkova Bay (KB), Russia (68°35′N, 52°20′E, Fig. 2c). While these sites vastly differ in geographical position, they are all lowland sites in close proximity to the coast, which facilitates comparison. Data collection in the colonies took place in different years: Kongsfjorden (2000, 2001, 2003, 2005–2016); Nordenskiöldkysten (2004, 2010–2016), Kolokolkova Bay (2003–2009, 2014, 2015). From geographic positions of geese equipped with tracking devices (Tombre et al. 2017; Lameris et al. 2018), we determined staging sites close to the breeding grounds, which geese use prior to moving to the breeding colonies. Geese forage at these staging sites until conditions become suitable for laying in the breeding colonies (Hübner 2006; Lameris et al. 2018). We identified three proximate staging sites, which were closest to the breeding colonies (Hübner 2006), on Svalbard: Lognedalsflya (LF), Vårsolbukta (VB), Sarsøyra (SØ), (Fig. 2b); and two sites around the Kolokolkova Bay: Neruta river delta (ND) and Molotsnii river delta (MD) (Fig. 2c). We further identified another three southern staging sites on Svalbard, which were at a larger distance from the colonies on Svalbard: Hornsundneset (HN), Ralstrånda (RS), Daudmannsøyra (DØ).
We used the period between the moment of snowmelt (spring) and snowfall (autumn) as a measure of the length of the Arctic plant growth season. Snowmelt is an important driver of the timing of reproduction (Madsen et al. 2007) through its effect on the phenology of Arctic plants (Prop and de Vries 1993; Livensperger et al. 2016). Snowfall in autumn, on the other hand, puts a rigid end to feeding opportunities and thus to the breeding season. We estimated daily snow cover (as percentage cover over total area) for all study sites and years (see Table S1 for an overview of sites) for the period with sufficient daylight (26 February–30 September) using satellite images of the MODIS snow cover product (MOD10a2 version 6, Hall et al. 2006). As the areas for the breeding colonies KF and NSK were too small to determine snow cover, we instead chose the nearest coastal tundra which geese used for foraging during egg laying (Prop and de Vries 1993). To limit the effects of clouds obscuring the image, composite satellite images were generated over 8 days. Any composite images with a cloud cover exceeding 25% were excluded. A pixel (500 m resolution) was assigned as snow when classified as snow at least once during an 8-day period. With a spatial overlay of the breeding areas with the MODIS images, the number of pixels classified as snow within the study site could be retrieved. From the number of snow pixels and the total number of pixels in the breeding area, we calculated the percentage of snow cover. We linearly interpolated between values from composite images to attain a daily percentage of snow cover. From the snow cover data, we extracted the date of snowmelt, which we defined as the first day of the season at which snow cover was less than 50% (a measure which correlates with date of peak food quality; Lameris et al. 2018). Similarly, we calculated the moment of snowfall as the last day of the season at which snow cover was less than 30%. We chose this cut-off value for snowfall as higher levels of snow cover were not always reached before 30 September. We calculated season length at the breeding colony as the interval between dates of snowmelt and snowfall.
Per nest, we determined egg-laying date (date when first egg was laid), clutch size (number of eggs laid), number of hatchlings (number of eggs which successfully hatched) and nest success (whether a nest produced hatchlings) for as many nests as possible (Table S2). The parameters and the precise methodologies varied among breeding colonies. We collected data on egg-laying dates, number of hatchlings and nest fate in all colonies, and data on clutch size in KB and KF colonies. In addition, we recorded the number of nests in the colonies.
In the KB colony, we systematically searched for nests and checked nests every 2–3 days during the laying and early incubation period (late May–late June). Eggs were marked and the number of eggs was recorded at every visit. In the early and mid-incubation period, we determined clutch size as the total number of eggs in a nest when encountered with the same number of eggs during two subsequent visits. We excluded nests in which egg dumping was evident (more eggs per interval than expected or additional eggs after clutch completion). We visited nests during hatch (mid-June–late July) every 2 days to estimate date of hatch and record nesting success and number of hatchlings. We recorded nest fate as successful, predated, flooded or abandoned. A nest was considered successful when at least one chick had hatched, which we determined either by presence of hatchlings at the nest or presence of egg membranes and trampled nest rim (Davis et al. 1998). Empty nests and nests containing eggshells without membranes were considered as predated, or considered flooded when the nest was partly under water. Nests encountered after the laying period containing cold eggs and without nest owners present were considered abandoned. We recorded the number of hatchlings when (1) at least 50% of the eggs were in the process of hatching (cracks, hatching or hatched chicks), or (2) less than 50% of the eggs had successfully hatched (thus goslings present) and other eggs did not show signs of hatching. For the number of hatchlings, we assumed all eggs with signs of hatching to produce hatchlings. Hatching success was calculated as the number of eggs hatched divided by clutch size. To minimize disturbance, not all successful nests were visited at hatch. The total number of nests found in the colony during the study period was recorded as a measure of nesting propensity.
In the KF colony, the same methods were applied, except that nests were only visited from the early incubation period onwards, and not in the laying period.
At the NSK colony, we observed the goose colony on an offshore island from an observation tower on the mainland, 200 m away from the colony. Nests were monitored 6–16 h/day during the period that nesting geese were present. Nests in view of the tower were mapped on high-resolution images of the island, which enabled us to assess the breeding history of individual birds by visual observation from laying until hatching. 30–60% of the pairs was recognizable by coded leg rings (either one or both partners carrying a ring). As we did not find a difference in any of the parameters estimated between marked and unmarked pairs, all pairs were used in subsequent analyses. To avoid disturbance, the island was not visited during the breeding period and, therefore, clutch sizes were not determined. Nest fate was established from direct observations, and rated as successful (at least one gosling was seen at the nest and no predation of eggs or goslings was observed), predated (eggs or chicks were taken by a predator, most often polar bears Ursus maritimus), or abandoned (nest owners abandoned the nest territory before the eggs hatched and prior to any predation event taking place—after which the eggs were usually taken by glaucous gull Larus hyperboreus). The number of hatchlings was recorded by visual observation of nests that successfully hatched. The first day that goslings were seen at the nest rim was taken as the date of hatch. The total number of nests was recorded for every year.
For KB and NSK, we used nest fate to calculate the nesting success as the proportion of initiated nests that successfully hatched per year. As nests were not observed during the entire incubation period in the KF colony, we did not calculate nesting success for KF. To combine nesting success and number of hatchlings into a single measure of reproductive success, we calculated the total number of expected hatchlings per nest, per laying date, year and colony, as the product of (1) nesting success and (2) average number of hatchlings in successful nests.
Methods to determine the date of egg laying differed among study sites. In KB, laying date was estimated by back calculation for clutches found during egg laying, assuming a laying interval of 33 h, as follows: day of discovery when one egg was found; day of discovery minus 1 at two eggs; day of discovery minus 3 at three eggs; day of discovery minus 4 at four eggs (van der Jeugd et al. 2009). Both in KB and KF, laying dates were also back calculated from hatch date. Hatch date was estimated for clutches found in the process of hatching as follows: date of observation was taken when the nest contained at least one egg with holes, a hatching chick or a wet chick; 1 day was subtracted from the date of observation when all chicks were fluffy and dry; 1 day was added to the date when the nest contained only eggs with cracks. For back calculation, we assumed a period of 29 days between laying date and hatch date (as derived from 573 nests in the KB colony between 2005 and 2015 for which both lay and hatch date were determined), which is similar to results from NSK (30 days between laying date and hatch date, derived from 99 nests for 2010–2016). In NSK, laying date was estimated as the first day during which a pair occupied a territory. Territories that were occupied for only 1 day were not considered in analyses.
Site-specific approaches in collecting data might affect the potential to make comparisons between study sites. In KB and NSK, where we used back-calculated as well as observed laying dates, the close proximity of the period between laying dates and hatch dates (see above) gives us reason to believe that these methods are comparable. By back calculating laying dates from hatching dates as done for the KF colony, we did not take into account the laying dates of nests which did not survive until hatch. However, we found no reason to suspect that this affected estimates of laying dates considerably, as extensive nest searches in the colony throughout the incubation period indicated that only few nests were lost (7.5% on average).
We tested relationships between date of snowmelt, laying dates and reproductive success by linear models in R 3.5.1 (R Development Core Team 2018), using the package “lme4” (Bates et al. 2018). We added year and/or study site (all sites where we measured snow cover) as random factors to account for either the different years during which data were collected when a trend over years was not of interest, or to account for data from different study sites when the specific sites were not of interest. Candidate models were constructed from all possible combinations of predictor variables, including interactions which were considered ecologically meaningful. All models were compared using Akaike’s information criterion corrected for small sample sizes (AICc; Burnham and Anderson 2004) and we chose the model with the lowest AICc value as our final model. Models within 2 ΔAICc of the final model were considered as competitive as long as these did not contain extra, potentially uninformative, parameters in comparison to the final model (Arnold 2010). Model-averaged parameter estimates were obtained by the package MuMln (Bartoń 2018). Support of the selected model (or models) relative to next best model was calculated from the ratio of model weights (Burnham et al. 2011). Besides predictor variables relating to snow cover, study year and the fitness components (clutch size, number of hatchlings and nesting success), we used predictor variables which separated high- and low-Arctic sites and staging and breeding sites, including ‘area’ (high or low Arctic), ‘site’ (all sites from which we gathered data on snow cover), ‘site type’ (southern staging sites/proximate staging sites/breeding colonies) and ‘colonies’ (the three study colonies).
First, to analyse if the snow-free period differed between the high and the low Arctic and among years, we ran linear mixed effect models (LMMs) with date of snowmelt/snowfall/season length as a response variable, year as fixed factor and area as fixed covariate, with site as a random factor. To test whether the snow-free period differed between breeding and staging sites, we ran LMMs with date of snowmelt as a response variable and site type and area as fixed factors, and with site and year as random factors.
Second, to analyse whether laying dates differed among years and between colonies, we used a linear regression model (LM) with average yearly laying dates as a response variable, and year and colony as fixed factors. To analyse how laying dates were affected by date of snowmelt, we ran LMMs with yearly average laying date as a response variable, date of snowmelt (in colonies and at proximate staging sites) and colony as fixed effects, and year as random factor. We tested whether the difference between laying dates and date of snowmelt at proximate staging sites differed between colonies by running an LMM with the difference in days as a response variable, colony as a fixed factor, and year as random factor.
Third, we aimed to analyse the association between laying date and date of snowmelt with fitness components. We ran generalized linear regression models (GLMs) with a Poisson distribution for clutch size and number of hatchlings as response variables, and GLMs with a binomial distribution with a logit link function for hatching success as a response variable. In these GLMs, we included colony as fixed factor and either date of snowmelt or laying date as fixed covariates. We tested the effects of these variables in separate models as the variables were highly correlated. A year effect was accounted for by date of snowmelt, and therefore, year was not included in the analyses as an additional covariate. We tested the association between clutch size and number of hatchlings in an LMM with year and site as random factors. Furthermore, we ran GLMs with a binomial distribution and a logit link function with nesting success as response variable. We included either laying date and laying date squared or date of snowmelt as predictor variables. For NSK, we excluded the years 2012 and 2014, when nest success was 0. We ran LMMs with total expected number of hatchlings as response variable, included laying date and laying date squared as response variables. We ran a similar analysis for total expected number of hatchlings in a GLM per year to retrieve slopes per year per site. We tested the association between number of nests and date of snowmelt in an LM, including colony as a fixed factor.