Fig. 2 Schematic of analytical framework and study glossary. Steps A-C show how a refined foraging circle (green) can be predicted for a seabird colony (yellow triangle). If tracking data originates from the colony then this is used to train a colony-specific model to predict habitat suitability and to inform a colony foraging circle; if not, habitat suitability is predicted by a multi-colony model and a global foraging circle is applied. Methods of refinement apply to both colony foraging circles and global foraging circles.

A framework integrating the foraging radius approach with ecological niche models

We integrated models of seabird foraging niche with the foraging radius approach by excluding predicted unsuitable foraging habitat within foraging radius circles to produce “refined foraging circles”. Taking a precautionary approach to minimise erroneously excluding good foraging habitat from within foraging circles, we specified that the area of unsuitable foraging habitat excluded within circles was dependent upon ENM transferability, which we term “transferability-supported refinement”. This allowed greater foraging circle refinement with more transferable models. The first step of our framework is to predict foraging habitat suitability and clip it to within the foraging circle of a colony of interest, limiting the foraging habitat available to the population. This allows us to derive percentiles of habitat suitability within the foraging circle which are mapped as contours. The second step scales ENM transferability values onto the percentiles of predicted foraging habitat suitability, delineating a refined foraging circle from habitat suitability values equal or above to the selected percentile. We scale transferability to percentiles ((AUC-0.5)/(0.9-0.5))*(0.9-0), such that AUC ≤ 0.5 (prediction no better than random) takes the 0th percentile of foraging habitat suitability values, thus defaulting to the unrefined foraging circle. Higher AUC values take higher percentiles up until AUC ≥ 0.9, which is set to an upper limit of the 90th percentile (a threshold proposed for translating seabird ENM predictions into marine Important Bird Areas by Dias et al. (2019b), thus allowing models with excellent transferability (AUC ≥ 0.9) to delineate refined foraging circles with the top 10% of foraging habitat suitability values.
Transferability-supported refinement simply offers a suggested percentile of habitat suitability with which to refine foraging circles, but planners may wish to select a different percentile to create a smaller or larger refined foraging circle. We term this “area-limited refinement” and anticipate that it may be necessary to use when transferability-supported refinement produces refined foraging circles that are still considered too large for area-based management.

Framework validation and attributing refinement confidence

To validate our framework, we simulated increasing foraging circle refinement at each colony in our global tracking dataset to estimate “refinement confidence”: the probability of known foraging areas (50% UDs) being included in refined foraging circles. The results allowed us to estimate how well unrefined global foraging circles capture known foraging areas and model the rate this inclusion declines when refining for different transferability ENMs. For each modelled species and colony (Table 1), we predicted foraging habitat suitability using the corresponding multi-colony model, trained on all colonies except the colony being predicted to. We predicted models to each colony using the same oceanographic variables and same month used in model training, assessed transferability using the multi-colony leave group out cross validation AUC value, and mapped foraging habitat suitability predictions onto a raster with two km cell size. At each colony, we refined the global foraging circle using the percentile selected by the model transferability AUC value, and also simulated refinement increasing from the 0th percentile (unrefined foraging circle) to the 90th percentile at five percentile intervals. Refined foraging circle polygons were obtained by binarizing foraging habitat suitability rasters with the specified percentile value, and tidied with R package smoothr (0.1.1) (Strimas-Mackey 2021) to remove small ‘crumbs’ and holes in polygons. In each iteration, we calculated the percentage of known foraging areas tracked from the colony that were included in the refined foraging circle. We fitted logistic regression models with percentage of known foraging area inclusion as the response variable and foraging habitat suitability percentile and its interaction with model transferability as explanatory variables, allowing the slope to vary within a species-colony random effect. Logistic regression models were fitted for each modelled species separately, and all combined, using R package lme4 (1.1-21) (Bates et al. 2015); random effects were dropped from TERN and SOTE regression models due to too few colonies.
The intercept and slopes of logistic regression models allowed refinement confidence to be predicted given model transferability and the percentile of habitat suitability selected for refined foraging circle delineation. This firstly allowed us to validate whether the AUC-percentile scaling was appropriate, under transferability-supported refinement we would expect to see high refinement confidence for all refined foraging circles, regardless of transferability (poor models refine a little, excellent models refine a lot; all retain known foraging areas). Secondly, logistic regression model coefficients allowed us to explore the trade-off between refined foraging circle size and refinement confidence as coefficients predict refinement confidence for any given size of refined foraging circle. When refining a foraging circle for species x at untracked colony y, its’ size can be suggested from transferability-supported refinement (using the multi-colony global transferability value) or specified by area-limited refinement. Having the associated refinement confidence predicted from logistic regression model coefficients for refined foraging circles of different sizes allowed us to judge the amount of refinement most appropriate during our Great Barrier Reef case study.

Foraging circle refinement on the Great Barrier Reef

To demonstrate the application of our framework, we refine foraging circles for the breeding seabird community of the Great Barrier Reef. From these we identify three networks of candidate MPAs (unrefined foraging circles, transferability-supported refined foraging circles, and area-limited refined foraging circles) to demonstrate the trade-off between total area and refinement confidence. For species without tracking data on the GBR, we predicted habitat suitability using multi-colony models. The remaining species had tracking data from one or two breeding sites on the GBR (Fig. 1). For these species, habitat suitability was predicted by colony-specific models to corresponding tracked colonies and neighbouring colonies within the same designated breeding site (see supporting information), and multi-colony models to the remaining area. We predicted models upon annual averages of monthly dynamic oceanographic variables, as the modelled species breed year-round across the GBR (but most show seasonal breeding peaks), with the exception of wedge-tailed shearwaters for which oceanographic variables were averaged over their distinct breeding season (December-April). Predictions were mapped across the GBR and Coral Sea using a raster with two km cell size. We placed foraging circles around breeding sites from significant seabird areas on the GBR. Significant seabird areas were either designated as internationally-recognised Key Biodiversity Areas (KBAs) or possessed regionally significant breeding populations of one or more species (see supporting information). For the seabird species listed at each breeding site we applied the corresponding global foraging circle, unless the breeding site was part of a significant seabird area with a tracked colony, in which case the colony foraging circle was applied.
Following our framework, we conducted transferability-supported refinement of global foraging circles using multi-colony model inter-colony transferability values and colony foraging circles using colony-specific model local transferability values. To investigate the effect of further refinement, we conducted area-limited refinement on any refined foraging circle produced by transferability-supported refinement that was over 100,000 km2, until all refined foraging circles fell below 100,000 km2(representing a maximum layer size in a hypothetical conservation planning exercise). The refinement confidence in colony-specific models was obtained using the inclusion of known foraging areas (50% UD) from tracked GBR breeding sites within refined foraging circles while the refinement confidence in multi-colony models was determined using the logistic regression model coefficients from our global framework validation exercise. To compare differences between different refinement approaches, we created three networks of candidate MPAs from unrefined foraging circles, transferability-supported foraging circle refinement, and area-limited foraging circle refinement (<100,000 km2). We summarised differences between networks by comparing the total foraging area required for the GBR breeding community, dissolving overlapping refined foraging circles shared by multiple breeding sites, and by comparing refinement confidence under each approach. Finally, we selected the most appropriate refinement approach for each species and merged results to create the most suitable network of candidate MPAs for seabirds on the GBR.