Seasonal migration has evolved across taxa and encompasses a multitude of features, many of which vary between species, between and within populations, and even within individuals. One feature of migration that appears especially variable within individuals is the route taken to reach a destination, even when the destination itself is not variable at this level. To investigate why, we analysed the geolocator tracks describing 192 post-breeding migratory journeys of 84 common terns (Sterna hirundo), as well as 149 pre-breeding migratory journeys of 75 of these birds. We found little within-individual spatial consistency in migration routes across years, irrespective of season or sex. Instead, individuals departing during the same time window took similar migration routes, which, during pre-breeding migration, when birds predominantly encountered headwinds, were associated with minimised headwind exposure. We therefore suggest that the individual routes of this long-distance migratory seabird can be flexibly adjusted to environmental variation, which is likely to be adaptive.

Between mid-May and early July 2016 – 2020, we caught 84 incubating common terns a total of 229 times with a drop trap, on average 16 ± 4.3 SD days after the first egg in their nest was laid. These birds were then equipped with a geolocator attached to the leg using a 10 mm aluminium ring [Kürten et al. 2019]. The geolocators were set to sample ambient light intensity every minute, with the maximum light intensity being stored every five minutes (mode 10). The total mass of the geolocator, ring and glue was 1.6 g, which equalled 1.2% ± 0.1 SD of the body mass of the birds at capture (mean body mass: 129.1 g ± 8.1 SD). In the breeding seasons of 2017 – 2021, we re-trapped the 84 individuals equipped with a geolocator that returned (i.e. 86%) a total of 198 times, and recovered 192 devices with light-level data (i.e. 97%). As such, we obtained data for 192 post-breeding migratory journeys performed by 84 individual birds (41 males and 43 females), as well as 149 pre-breeding migratory journeys performed by 75 of these birds.

To analyse the light-level data, we used the function “preprocessLight” of the R package “BAStag” [Wotherspoon et al. 2016] and a threshold of 1.5 to identify twilights. If visual inspection suggested extreme outliers (>30 min difference with the previous and subsequent twilight), we either adjusted the outlying twilight or excluded it (c. 1% and < 1%, respectively; as calculated based on the light-level datasets of 10 randomly selected individuals).

All geolocators were calibrated ‘on-bird’ while the birds were at the Banter See colony (i.e. for multiple weeks). The start and end of the calibration periods were individually determined by visual inspection of the calibration slopes for sunsets and sunrises plotted by the “plot_slopes_by_location” function of the R package “FLightR” (version 0.5.0, [Rakhimberdiev et al. 2017]). As long as the bird is staying in the calibration location, these slopes vary very little [Rakhimberdiev et al. 2017], so that the first calibration period ended with a visual detection of a change in the slope (i.e. when the bird was assumed to have departed). For geolocators that where still working at recapture, a second calibration period was set to start when the slope did not show variation anymore (i.e. when the bird had arrived).

For the subsequent movement analyses, we set the mean flight distance between two twilights to 1200 km, i.e. 24 hours x 50 km/h [Bruderer & Boldt 2001], while allowing for ± 300 km and ran the particle filter with 1 million particles to optimise the track of each individual and to minimise its uncertainty [Rakhimberdiev et al. 2015]. The uncertainty of location estimates obtained with “FLightR” is approximately 250 km per location (dependent on shading caused by the species’ behaviour), which is substantially less than that of estimates obtained with more traditional threshold methods [Rakhimberdiev et al. 2017, Halpin et al. 2021]. We deem this uncertainty to be unlikely to impact our analyses substantially, since all tracks are expected to be equally affected, common terns migrate a distance (5,000 – 11,000 km; [Kürten et a. 2022]) that is relatively large compared to the uncertainty in the location estimate, and our large sample size (n = 192 post-breeding and 149 pre-breeding migratory journeys) is securing statistical power.

Finally, we applied the stationary.migration.summary” function of “FLightR” with a “min.stay” of ten twilights (i.e. 5 days) and a “prob.cutoff” set to 0.4 (for detailed testing and explanation of these settings, see [Kürten et al. 2022]) to derive the longitude and latitude of each non-breeding residence area, as well as the dates of arrival to, and departure from, the breeding colony and each non-breeding residence area for each track from each bird. For 192 tracks from 84 individuals, we obtained a single non-breeding residence area, whereas for 7 tracks of 6 individuals we obtained two. The post- and pre-breeding migration routes were obtained from the daily location estimates between the departure dates from, and arrival dates to, the breeding colony and first and last non-breeding residence area, respectively.