Tracking data processing

Tracking datasets were speed filtered (removal of points > 90 km/h; (Mendez et al. 2017), and linearly interpolated using theAdehabitatLT (0.3.24) R package (Calenge 2011). Due to differences in temporal resolution of different datasets we interpolated each dataset to either 1, 2, 3, 5, 10, 15, or 20 minute resolutions. We split individual foraging trips from multi-day tracks using thetrack2kba (1.0.0) R package (Beal et al. 2021), removing small foraging trips within 4 km of the colony and under 1 hr in duration. We also manually removed trips that spent too much time away from the colony, indicating breeding failure. Upper trip duration limits were set at five days for boobies (Mendez et al. 2017), 12 days for frigatebirds (Mott et al. 2016) and tropicbirds, 14 days for wedge-tailed shearwater long trips (McDuie et al. 2015) and sooty terns (Neumann et al. 2018), three days for wedge-tailed shearwater short trips (Weimerskirch et al. 2020) and two days for noddies and terns (from inspection of data).

Foraging radii

We estimated radii and mapped foraging circles for each modelled species to predict likely foraging range around untracked colonies (Birdlife International 2010; Thaxter et al. 2012). We first obtained the maximum distance from the colony observed across all the foraging trips made by birds from each tracked colony. We then took the average of these colony-specific maximum distances to generate a ‘mean maximum foraging radius’ for each modelled species (Thaxter et al. 2012). To provide lower and upper extremes for each modelled species, we also present the minimum and maximum of colony-specific maximum distances observed.

Ecological niche modelling

To model the foraging niche of each modelled species, we assumed a binomial response comparing the oceanographic covariates of known foraging areas (1) against the oceanographic covariates of accessible habitat (0). To identify known foraging areas for each dataset we performed location-based kernel density analyses with a 1 km grid on a subset of tracking datapoints identified as representing foraging behaviour (Miller et al. 2017; see supporting information for further details), and treated all grid points within the 50% utilization distribution (UD) as ‘presence’ points in the model. The accessible habitat for each colony was defined as the convex hull containing all tracking locations (all behaviours included) and curtailed to marine regions. Areas inside convex hulls were sampled using pseudo-absence datapoints distributed randomly but weighted by inverse distance to the colony (to constrain access to foraging habitat by central-place foraging seabirds). Pseudo-absences were created at a rate of 3:1 to presences (Wakefield et al. 2011) and given the same timestamp as their respective presences for dynamic covariate extraction. To account for the potential influence of random sampling of pseudo-absences on model stability (Barbet-Massin et al. 2012), we repeated the random selection of pseudo-absences five times, generating five replicate presence and pseudo-absence datasets per original dataset. Each of the five replicates was modelled separately and then averaged together for model validation and prediction.
Although tropical seabird prey opportunities are patchily distributed and ephemeral in nature (Weimerskirch 2007), their location and availability are governed by physical ocean processes at broader spatial scales (10-100kms; Wakefield et al. 2009). Ocean covariates for modelling were created to capture broad-scale ocean features representing attractive and/or reliable locations for tropical seabird foraging (Table 2). For dynamic covariates, chlorophyll concentration, sea surface temperature and frontal activity, we created a long-term (~10 year) mean for each month of the year to model whether birds target specific covariate values, and also created the long-term standard deviation over 12 months of the year to describe whether birds target areas that are temporally dynamic/homogenous. The month in which each tracking dataset commenced was used to select the monthly dynamic covariate layer to extract values from. The use of long-term averages in models means that the oceanographic conditions fitted against known foraging areas of tracked birds do not reflect the concurrent oceanography. Rather, they characterise persistent oceanographic features in the seascape. The advantage of this approach is that a global layer for each oceanographic variable can be created, meaning subsequent predictions of seabird foraging at each colony are all built on the same, standardised, comparable oceanography. This is particularly important for assessing model transferability between colonies. Models also included static covariates of bathymetry and seabed slope to allow them to describe the importance of geographical features such as reefs, shelves and seamounts for foraging. Following the seabird foraging niche transferability study of Péron et al. (2018), we did not include a distance-to-colony variable because it is explicitly linked to colony-specific demographic information, such as population size (Lewis et al. 2001), which we lacked for training and test colonies. All data handling and statistical analyses were performed in the statistical software environment R version 3.5.1 (R Core Team 2020). For more detail on tracking and oceanographic data processing see supporting information.
Table 2 Oceanographic data sourced for modelling.