Fig. 3 Model validation and transferability for modelled species (A-J).
Matrix values describe predictive accuracy via area under the receiver
operator characteristic curve (AUC); AUC values (0-1) have been
multiplied by 100 to save space. Reading by row shows how well each
model trained on colony-specific data predicts foraging habitat
suitability at all other colonies, reading by column shows how well
foraging habitat suitability at each colony is predicted by models
trained with data from all other colonies. Row names denote each
colony-specific model and the multi-colony model. The ‘MEAN’ column
gives global transferability of each colony-specific model and the
multi-colony model by averaging the predictive ability across all
colonies in the row (excluding the self-prediction value in
colony-specific models). Local prediction is compared between
colony-specific model self-prediction (diagonals) and multi-colony
models (bottom row) by colour of AUC value: green shows model is better
than the other; red shows model is worse than the other; black shows
models are comparable (AUC values within 0.05 of each other). AUC value
background colour denotes performance: grey (no better than chance/very
poor); yellow (poor); light orange (moderate); dark orange (good); red
(excellent). Dendrograms show clustering of pairwise colony AUC values;
closely clustered colonies predict one another better. Family group
models (B, C & G) have species codes added to colony names: CRTE =
crested tern, CATE = caspian tern, ROTE = royal tern, LENO = lesser
noddy, BLNO = black noddy, BRNO = brown noddy, GRFR = great frigatebird,
LEFR = lesser frigatebird, MAFR = magnificent frigatebird. All colonies
in the family group model for tropicbirds (H) represent red-billed
tropicbird, with the exception of the Ashmore Reef (Australia) colony,
which is red-tailed tropicbird (denoted with *).
Tropical seabird foraging
radii
When considering all species and colonies, foraging radii ranged from
28.4 to 1166.2 km (Table 3). By increasing foraging range, modelled
species were generally ordered: terns, brown booby, noddies, masked
booby, wedge-tailed shearwater short trip, red-footed booby,
tropicbirds, sooty tern, frigatebirds, wedge-tailed shearwater long
trip. We found that the global foraging circle for each species
contained a high percentage (89-100%) of known foraging areas from each
tracked colony (Table 3).
Validating foraging circle
refinement
We found an average of 89% of known foraging areas were included within
refined foraging circles when using transferability-supported refinement
of global foraging circles around each colony from our tracking dataset
(Fig. 4). Our AUC-percentile scaling was effective at limiting the
exclusion of erroneously predicted foraging habitat in poorer
transferability models giving 80-99% known foraging area inclusion when
the 45th percentile of habitat suitability or below,
was used to delimit refined foraging circles. For refined foraging
circles delimited with habitat suitability above the
45th percentile, our scaling maintained known foraging
area inclusion at ~80%, showing that greater foraging
circle refinement was effectively offset by the more accurate prediction
of foraging habitat by higher transferability models.
Simulating increasing refinement of global foraging circles around each
colony from our tracking dataset demonstrated that in the absence of
transferability-supported refinement, refined foraging circles delimited
by less transferable models always had a lower probability of including
known foraging areas than those delimited by more transferable models
(Interaction between percentile of foraging habitat suitability and
model transferability: χ21 = 100.1, p
< 0.001; Fig. 4). Simulation results for individual modelled
species demonstrated the same pattern but there were differences between
species due to ecology and sampling effort (Table 4, visualised in
supporting information). Inclusion of known foraging areas in global
foraging circles set the baseline probability for inclusion prior to
refinement and differed between species (e.g. high in frigatebirds, low
in sooty terns – despite similar foraging ranges; Tables 3 & 4). The
steepest declines in known foraging area inclusion with foraging circle
refinement were seen in masked and brown boobies and tropicbirds,
whereas terns retained the highest inclusion of known foraging areas
under refinement because their multi-colony model had good global
transferability. Improving model transferability gave the greatest boost
to inclusion of known foraging areas within refined foraging circles in
tropicbirds, brown booby, sooty tern and masked booby, whereas
frigatebirds known foraging area inclusion was least affected by model
transferability (Table 4). The noddies model did not include the
interaction term with transferability as all multi-colony model
predictions were below 0.5 AUC.