Ecological Realism
Similar to their predictive skill, all models’ ability to generate
ecologically realistic environmental relationships and spatial
predictions varied primarily by spatial dependence treatments. Response
curves indicated that each model captured similar relationships between
albacore presence observations and the environmental variables. However,
spatially explicit models (i.e., GF and iSDM) demonstrated a broader
relationship with MLD compared to the HE model. For example, the HE
model response curves suggested albacore suitability was highest at
shallow MLDs (~ < 10 meters; Figure 5). This
relationship closely resembled the underlying MLD distributions of the
logbook data (Figure S1), which suggested significantly deeper MLDs
relative to the archival tag data (mean MLD: 32.7 ± 21.9 and 13 ± 4.03
for archival tag and logbook data, respectively). In contrast, the GF
and iSDM models suggested high albacore habitat suitability for shallow
MLDs to ~100 meters before declining with depth (Figure
5). Additionally, all models showed that response curves (i.e., from the
training period) fully covered the conditions during the forecasting
period, though the distributions for each environmental variable
narrowed during this time (Figure 5). Despite this coverage, the
distributions of environmental values differed between the training and
forecasting periods. For example, SST was warmer, bathymetry shallower,
and MLD nearer to the surface during the forecasting period, indicating
that albacore inhabited warmer, shallower waters during MHW conditions
(Figure 5).
Spatial predictions were also assessed for ecological realism. During
the forecast period under MHW conditions, the models generally exhibited
similar spatial patterns in predicted habitat suitability, though
regional differences were apparent (Figure 6). All models identified
coastal waters—particularly off the coasts of Washington, Oregon, and
northern Baja California—and the North Pacific Transition Zone as
highly suitable habitat. However, the HE model produced the least
ecologically realistic predictions, as it forecasted higher habitat
suitability across the study area compared to the spatially explicit
models. Notably, the HE model predicted extensive areas of suitable
habitat for albacore tuna in areas characterized by warm (positive) SST
anomalies offshore of southern California, whereas the GF and iSDM
models demonstrated a discernible discontinuity in suitability within
this highly anomalous region (Figure 6). This overestimation in the HE
model is further reflected in the distribution of predicted suitability
values, where it overpredicted habitat suitability at pseudo-absence
locations. In contrast, the GF and iSDM model predictions more
accurately differentiated between presences and pseudoabsences,
resulting in a more balanced, and likely realistic, distribution of
suitability values (Figure 6). The primary differences between the
spatially explicit models were in the iSDM predictions being more
closely aligned with albacore occurrences in coastal Canada and absences
off central California, whereas these regions were connected by a
coastal band of apparent high suitability in the GF predictions. In
addition, while both GF and iSDMs identified the waters off Baja
California as suitable, the degree of predicted suitability was
significantly higher for the GF model in this region.