Discussion

Relying on the most comprehensive tropical seabird tracking dataset to date, we investigated whether ENMs can transfer (predict) foraging niches of breeding tropical seabirds between global colonies, and whether ENMs can be combined with the foraging radius approach (Thaxter et al. 2012) to refine foraging circles around breeding colonies. We found little ability to generalise and transfer ENM predictions across all colonies for any tropical seabird species. However, we frequently observed clusters of colonies that predicted well to one another, but poorly to other colony clusters. Despite the limited ability of ENMs to predict foraging niche in new regions, they were able to refine foraging circles by excluding predicted areas of unsuitable foraging habitat. We found inclusion of known foraging areas was almost certain in unrefined foraging circles, and remained high when foraging circle refinement was specified by model transferability, where neither greater refinement from high transferability models nor minor refinement from poor transferability models erroneously excluded important foraging habitat. When applied to the Great Barrier Reef, this framework was able to reduce, with confidence, the area required to protect foraging resources of the breeding seabird community.

ENM transferability

Limited ENM transferability can be caused by differences in the range and/or combinations of environmental variables between training and test datasets (extrapolation); poor description of underlying processes by explanatory variables in models (misspecification); and differences in pressures (e.g. competition, predation, local marine productivity etc) between training and test populations (local adaptation) (Randin et al. 2006, M. McPherson & Jetz 2007, Torres et al. 2015, Péron et al. 2018). Extrapolation likely impacted transferability of colony-specific models, as they were more likely to encounter novel environmental values when predicting to test colonies. By contrast, training of multi-colony models across numerous colony-specific environmental ranges reduced the likelihood of extrapolation when predicting to test colonies, and could explain their slightly better global transferability compared to colony-specific models (Table 3).
Misspecification of models could have contributed to limited global transferability, as we tried to predict seabird foraging with long-term averages of oceanographic variables. Our goal here was to characterise seabird foraging associations with persistent oceanographic features in the seascape rather than ephemeral ocean phenomena, as predictive maps representative across space and time are likely to be the most pertinent for conservation scientists and managers (Guisan et al. 2013). The downside of this approach is that foraging areas selected by GPS-tracked birds do not reflect the concurrent oceanography in our models, which could miss foraging association with fine scale dynamic features (e.g. intra-seasonal upwellings), known to attract tropical seabirds (Kai et al. 2009, Miller et al. 2018). Model misspecification is supported by a general inability of colony-specific models to predict their own foraging niche (diagonal values in Fig. 3) and attributed to using long-term oceanographic variables because similar studies using high-resolution oceanographic variables produced good self-prediction (Péron et al. 2018). Future marine transferability studies should consider their study objective (Yates et al. 2018) (e.g. informing management or ecological understanding) and the dynamism of their ocean region (e.g. temperate shelf break vs pelagic frontal system) when selecting spatial and temporal scale of ocean covariates to use in ENMs.
We consider local adaptation a major cause of limited transferability in this study because multi-colony models did not provide a great improvement in transferability over the colony-specific model average. This finding suggests a limited ability of multi-colony models to generalise patterns in local adaptation of foraging niche from multiple colonies (Gilmour et al. 2018), and boost their transferability (Matthiopoulos et al. 2011). This could be explained by local adaptation causing homogenisation: when models fitted at large spatial scales average the responses of populations from contrasting habitats, and fail to capture and predict local extremes (hotspots and coldspots) (Paton & Matthiopoulos 2016). Model homogenisation likely affected all species known for local adaptation: frigatebirds (Mott et al. 2016); tropicbirds (Diop et al. 2018); wedge-tailed shearwater (Weimerskirch et al. 2020); and in particular boobies, known for their extreme foraging plasticity (Mendez et al. 2017, Gilmour et al. 2018).
We failed to explain inter-colony transferability with two factors describing local adaptation (geographic distance and oceanographic similarity; Redfern et al. 2017, Gilmour et al. 2018), suggesting colony demographic information could be an important missing factor (Paton & Matthiopoulos 2016). Breeding colony size is known to be a major driver of local adaptation in seabird foraging behaviours, where density dependent competition forces greater foraging ranges at larger colonies (Ashmole 1963, Lewis et al. 2001). Competition pressure may also come from neighbouring breeders, such that populations segregate foraging areas (Bolton et al. 2018), and from different species (Oppel et al. 2015, Mendez et al. 2017). However, including demographic covariates within ENMs requires complex modelling (Wakefield et al. 2017), which given their interaction in driving local adaptation may extrapolate poorly beyond the training region.
We modelled seabird foraging niches excluding a distance-to-colony variable (Péron et al. 2018), which is commonly included in seabird ENMs. Distance-to-colony is often the most important explanatory variable (Oppel et al. 2017, Miller et al. 2019) in seabird ENMs but encodes local adaptation as it is driven by colony-specific demographic information (Lewis et al. 2001). It is difficult to see how ENMs dominated by distance-to-colony could accurately predict foraging areas at unknown colonies unless training and test colonies had similar demographic pressures. Furthermore, if modelling multiple colonies together, model homogenisation would cause an average distance-to-colony to be predicted, potentially yielding the same result as the foraging radius approach (predicting a colony buffer). Our results show that modelling seabird foraging niches using long-term ocean variables alone yields generally poor transferability, and further study is needed to assess whether including distance-to-colony boosts transferability and generates spatial predictions useful for informing management.

Refining foraging circles

Minimising the cost of protected areas while maximising the confidence that they appropriately serve a species (e.g. contain core habitat) is a key goal of systematic conservation planning (Margules & Pressey 2000), and the refinement of seabird foraging circles presents a good example of this challenge. We demonstrate high confidence that known foraging areas are included within presented global foraging radii estimates (Table 3). However, foraging circles from these radii are too large to implement practical conservation measures across, particularly for wider ranging seabird species (Soanes et al. 2016, McGowan et al. 2017), for which an area-based conservation approach may not be efficient or desirable (Oppel et al. 2018). Although we acknowledge that cost and area are not analogous within conservation planning, smaller protected areas for seabirds reduce likelihood of conflict with other marine users (e.g. fishers) and require lower monitoring/policing effort (work hours, fuel etc) relative to larger areas. Using predicted habitat suitability to refine foraging circles provides a way of reducing protected area size, but the confidence in predictions must be considered so that conservation efforts are not allocated to the wrong areas. Our framework accounts for confidence by making the level of foraging circle refinement dependent upon ENM transferability. As we found that inclusion of known foraging areas remained high when foraging circles were refined using transferability-supported levels, we can be confident that neither greater refinement from high transferability models nor minor refinement from poor transferability models erroneously exclude important foraging habitat.
Even minor refinement of foraging circles adds information to marine spatial planning when multiple breeding seabirds are considered together. Overlaying unrefined foraging circles from multiple species just shows increasing overlap of concentric circles towards colonies, informing planners only that areas of sea surrounding the most seabird species-rich islands are the most important. These can be improved by weighting foraging circles with their breeding populations and distributing birds over an accessibility surface (inverse colony distance) within the circle (Critchley et al. 2018), but this still lacks ecological realism. The refinement of foraging circles with ENMs integrates habitat preferences into the planning process. For GBR seabirds, this reveals that areas offshore from colonies, particularly open sea adjacent to outer reefs, are likely foraging hotspots for multiple species (Fig. 6).
Our study of foraging circle refinement presents several tools for tropical seabird conservation, whose use we advocate in a hierarchical manner based on local data availability. Firstly, we present the most comprehensive collation of tropical seabird foraging radii to date. The estimates presented for each species can be used to represent foraging ranges for any population in the world without local tracking, and the colony-specific foraging ranges presented in the supporting information form a valuable resource for users interested in specific regions. Foraging radii alone have important applications, particularly the mean-maximum foraging range, such as assessing seabird population connectivity with planned offshore energy generation projects (Woodward et al. 2019). Secondly, where a candidate seabird protected area is required for a tracked colony, our framework can be followed to generate a colony foraging circle and refine it using a colony-specific model trained with the tracking data. Thirdly, where candidate seabird protected areas are required for untracked colonies and tracking data exist from several colonies in the same region, a global foraging circle can be estimated and refined with a multi-colony model trained on the regional tracking data. For species without any local tracking (frigatebirds, tropicbirds, red-footed booby, sooty tern and terns in our GBR study), we advise cautious application of a global multi-colony model. The generally poor transferability of global ENMs would only prescribe minor refinement of global foraging ranges, but we nonetheless advise local expert opinion or distribution data (e.g. at-sea surveys) should be used to verify that predicted ‘unsuitable’ habitat is indeed unsuitable. Refined foraging circles can be considered candidate MPAs for their respective seabird population. As distinct polygons they can be considered individually or overlapped, to identify multi-species foraging hotspots, in higher level marine spatial planning exercises, to ensure seabird representation in multi-taxa MPA delineation. It should be noted that our foraging circle refinement framework is ENM neutral, and users should select environmental covariates and model algorithms of their preference.

Advancing regional knowledge of seabird foraging areas

ENMs trained on GBR tracking allowed better refinement than globally trained models, demonstrating the value of local tracking data. A key recommendation from our study is collection of more seabird tracking data at regional level. If there was a representative (good coverage of species and sites) GBR database of seabird tracking information, there would be no need to apply model predictions from across the world to the GBR as we have done here. Coordination of a systematic regional seabird tracking campaign offers the most efficient solution to credibly identify known foraging areas for the GBR seabird community, as demonstrated by projects in the UK (FAME and STAR; Wakefield et al. 2017). Key breeding populations should be prioritized for tracking, but it is essential that colonies from the same species are tracked in different areas of the GBR. Our observation that foraging habitat suitable for brown boobies from Swain Reefs could not predict that of conspecifics from Raine Island (and vice versa; Fig. 6A) highlights the limits of model transferability within the same region. Furthermore, it may be necessary to investigate whether models of foraging niche are transferable between neighbouring colonies, in particular when partitioning of foraging areas between colonies is observed, as shown for wedge-tailed shearwaters breeding in New Caledonia (Weimerskirch et al. 2020). Selected colonies should be tracked during years of typical ocean conditions with a good sample of birds (>30; Soanes et al. 2013, Lascelles et al. 2016) to ensure the observed foraging niche is representative of the populations’ true foraging niche. A systematic seabird tracking campaign would dramatically reduce uncertainty in where seabirds forage on the GBR, enabling better focused management actions and inclusion of seabird foraging areas in higher level planning such as zoning of MPAs within the GBR. Nonetheless, the globally-informed predictions of habitat suitability and foraging radii presented here form the best working hypotheses of where seabirds forage on the GBR, and are a valuable starting point for management and protection of seabird foraging resources.