Manx shearwater (Puffinus puffinus) rafting behaviour revealed by GPS tracking and behavioural observations

Field methods: behavioural observations

We conducted field work on Skomer Island, a small island off the coast of West Wales, UK (51°44′08.9″N, 005°17′47.0″W). Behavioural observations were made of rafting Manx shearwaters surrounding the entire circumference of the island between August 5th and September 4th, 2016 from four observation stations (Skomer Head, South Haven, North Haven, Garland Stone; Fig. 1). These stations were selected because they avoided disturbing seabird burrows, offered full coverage of the island, and optimised rafting observations due to close proximity to the coastline and a clear field of view. The order of observation visits to each station was randomised to prevent systematic bias in the size of rafts and distance from shore due to the time of day. Behavioural observations were conducted by the same observer (Richards), which removed observer-related variability between estimates.

To record the size, bearing, and distance of each raft within a 2,000 m radius of the observation stations, each day, two hours before sunset, we used 8 × 42 binoculars, a magnetic compass, and inclinometer. The inclinometer was calibrated to zero using a spirit level, and the tripod holding the inclinometer was made level at each observation station using an inbuilt spirit level. Rafts observed at the 2,000 m boundary and beyond were excluded in our analysis because we expected higher error associated with the inclinometer and counting individuals. Daily weather data were from a weather station at Wooltack Point (51°44′11.3″N, 005°14′50.6″W), 3,400 m away from the centre of Skomer Island. Rafting observations were not carried out on nine days during the study period due to compromised visibility resulting from heavy precipitation, fog, or wind speeds exceeding 18 ms⁻¹.

Observational rafting distance and coordinates

The distance of each raft from the observation locations was calculated using the following basic trigonometric approach: (1)${\displaystyle Raftdistance=\frac{\left(C\pm T\right)}{D}\times H}$ where C = height of cliff at observation station (Garland Stone: 64.1 m; Skomer Head: 66.7 m; South Haven: 63.7 m; North Haven: 49.0 m), T = tide height relative to mean sea level, D = declination distance, and H = height of inclinometer (0.895 m). To determine the altitude and coordinates of the observation stations, values were taken from a GPS (Garmin etrex 10 software version 2.90). Tidal data were from Martin’s Haven (51°43′59.9″N, 005°15′00.0″W), 3,200 m away from the centre of Skomer Island. The full trigonometric approach is detailed in Equation S1.

Spherical trigonometry was used to calculate the rafts coordinates along a great circle given the bearing and distance from the observation stations. For statistical analysis, the distance of each raft from the nearest point of Skomer Island’s shoreline was estimated using the Haversine formula which calculates the great circle distance between two points.

Field methods: GPS tracking

Between 31st May and 21st August, 2016, adult Manx shearwaters (n = 83) were fitted with GPS loggers (Mobile Action iGot-U gt120), housed in heat shrink plastic and secured to small bunches of contour feathers on the back using strips of marine Tesa^® tape as in Guilford et al. (2008). GPS loggers weighed 17 g, comprising approximately 4.25% of a Manx shearwater’s typical body weight. GPS loggers were programmed to take location fixes every five minutes. Breeding birds were selected from specific burrows at a colony in North Haven on Skomer Island (51°44′14.0″N, 5°16′55.5″W). From each breeding pair, one bird was fitted with a GPS logger after its chick had been fed, or during partner changeovers for incubating birds. GPS loggers were removed the following night, or as soon as the individual returned to its burrow. Simultaneous deployments were avoided on pairs from the same nest. A maximum of 19 GPS loggers were simultaneously deployed during the study period. Handling time was kept to a minimum (normally less than 15 min). These GPS tracking methods have previously been shown not to have a measurable impact on breeding success at this study colony (Dean et al., 2012; Shoji et al., 2015).

This research was provided full approval by the University of Oxford Animal Welfare and Ethical Review Body and BTO Special Methods Technical Panel. Field experiments were approved by the Islands Conservation Advisory Committee (ICAC). A bird ringing permit was obtained from the British Trust for Ornithology (Permit Class and Number: / C / 5311).

Identifying rafting from GPS data

GPS tracking can further extend our understanding of rafting behaviour during the night and further afield, which is a limitation of visual observations. Consequently, the study area was extended from 2,000 m to 5,000 m to capture a wider image of rafting behaviour, whilst remaining focused near the island. To ensure behaviour at the colony and on land were not included in the analysis, all GPS fixes on land were removed from the dataset.

The Vincente ellipsoid function was selected to calculate the geodesic distance between GPS fixes. For each bird, the speed between GPS fixes was calculated as distance change over time. Erroneous locations were removed from the GPS tracks by applying a speed filter to omit all data points that exceeded 30 ms⁻¹. Since rafting and flying can be distinguished based on speed, we employed an objective and repeatable method for choosing a speed threshold that identifies the speed that, above which, a bird is more likely to be flying and below which, a bird is more likely to be rafting. To do this, we assumed that these two different behavioural states would each give rise to normally distributed speeds and thus, that the overall distribution of speeds would be a mixture of two normal distributions with a centre corresponding to each behavioural state. To parameterise these two normal distributions, we used an expectation maximisation algorithm, implemented using the ‘mixtools’ package in R and function normalmixEM with the number of components set at k = 2 (Benaglia et al., 2009), which iteratively fits two normal distributions to these emission data and finds the two means and standard deviations most likely to give rise to the observed data. Thus, from the 83 Manx shearwater GPS tracks, the speed at which a datum switched from being more likely to be attributed to the rafting centroid to being more likely to be attributed to flying was 2.5 ms⁻¹. Rafting was therefore defined as GPS fixes below a speed threshold of 2.5 ms⁻¹. All flying behaviour was filtered from the GPS tracks and only rafting behaviour was carried forward to further analyses. When mapped, the resulting rafting behaviours appear as close successive fixes showing a directional pattern following tidal or wind movements (Dean et al., 2012; Carter et al., 2016); Fig. 2).

Mapping, data conversion, and analysis

All analyses and mapping were carried out in the R 3.5.0 programming language and environment (R Core Team, 2018). Maps and figures were created with the packages ‘ggplot2’ (Wickham, 2016) and ‘ggmap’ (Kahle & Wickham, 2013). Several data conversions were required. To determine rafting direction, the bearing between each GPS fix was calculated using spherical trigonometry for each bird. All bearings were converted to radians and classified as circular data fitted to a Von Mises distribution. To explore the rafts’ response to wind direction, the rafting time and date were matched to the nearest time and date of the weather data from Wooltack Point.

Statistical tests

To test whether raft size changes with respect to wind speed within the behavioural observation data, a generalized least squares (GLS) model was constructed using the ‘nlme’ package and function gls (Pinheiro et al., 2018). Wind speed was included as a predictor (fixed effect), and raft size was log-transformed to meet the error structure assumption of homoscedasticity. Dates were converted to Julian days and temporal autocorrelation in raft size was modeled by fitting an ARMA process to the residuals. The number of auto-regressive parameters were specified as p = 4, and the ARMA was nested within each Julian day. Model results (based on the function summary.gls with default settings) are reported in Table 1, and the model structure and selection procedures are presented in the Code S1.

To test for a response in rafting direction (behavioural observations and GPS tracks) to wind direction, we used Jammalamadaka-Sarma circular correlation test using the package ‘circular’ and function cor.circular (Agostinelli & Lund, 2017). This test is a circular version of the Pearson’s product moment correlation and was selected because it allows the calculation of a correlation coefficient using angular variables.

To describe the trends in (1) the number of birds in nearshore waters through time and (2) raft distance from shore through time, we fitted general additive mixed models (GAMM) with Poisson distributions for count data. Bird ID was further included as a random effect for distance from shore through time. We used the package ‘mgcv’ and function gamm (Wood, 2017).

Rafting behaviour

From the behavioural observations, a total of 248 rafts were identified across 22 days, however, we removed 18 rafts because they fell outside the 2,000 m sampling boundary. Raft size was highly variable; the smallest raft consisted of 10 birds; the largest of 2,900 birds; and mean size was 336 ± 470 SD birds. A total of 77,390 individuals were counted over the duration of the study.

We found that Manx shearwaters rafted in all directions around Skomer Island (Figs. 1 and 3). Within the behavioural observation data, mean rafting distance from the island’s shoreline was 924 ± 315 SD m, and no rafts were located closer than 276 m; the result was a halo effect around the island with an absence of rafts within a 300–500 m boundary two hours before sunset (Fig. 1). The GPS data also revealed that Manx shearwaters raft further from land during the day resulting in a halo effect around the island two hours before sunset and after sunrise (Figs. 3A–3C). However, the halo effect disappeared at night as tracked birds rafted closer to the island (Fig. 3D). Rafting accounted for 85%, and flying for 15% of GPS tracked Manx shearwater behaviour within the 5,000 m radius of Skomer Island. Around the colony, a significant proportion, 93%, of rafting occurred during the night whilst only 7% occurred during the day.

Predictors of raft size

Wind speeds during the study period ranged from 2.1 ms⁻¹ to 17.6 ms⁻¹, and average wind speeds were 8 ms⁻¹. This is moderately representative of the season on Skomer Island because mean wind speed between April and September, 2016 was 7 ms⁻¹, and ranged from 0.3 ms⁻¹ to 20 ms⁻¹. More birds were present in rafts on windier days (Table 1). Residual autocorrelation was detected in the model of wind speed as a predictor of raft size. We accounted for this temporal autocorrelation with a lag of four days which significantly improved the model fit. When wind speeds were greater than 10 ms⁻¹, birds formed flocks of more than 10,000 individuals that failed to raft, or only settled for seconds. Birds were observed to be restless during periods of high wind, but were more settled in calm conditions.

Predictors of rafting locations

Wind direction was variable throughout the study period, and came from all directions except the east. There was no effect of wind direction on the rafting location of Manx shearwaters around Skomer Island within the behavioural observation data (Jammalamadaka-Sarma Circular Correlation, correlation coefficient = −0.015, test statistic = -0.228, p = 0.820). For the GPS data, there was a small negative effect of wind direction on rafting direction, (Jammalamadaka-Sarma Circular Correlation, correlation coefficient = -0.065, test statistic = −3.20, p = 0.001). During the night, the vast majority of GPS tacked birds rafted adjacent to the North Haven colony, located in the Northeast of Skomer (Fig. 3D).

Day-night trends in bird presence and rafting distance from shore

We found a non-linear trend in GPS tracked birds, where time of day affected the distance of rafts from the island and the number of birds in the nearshore waters. More birds arrived (Fig. 4A) and rafted closer to land (Fig. 4B) as the evening progressed, and as dawn approached birds moved further away (Fig. 4B) and left vicinity of the island (Fig. 4A).