Predicted habitat
The EXP approach selected the maximum number of optimal models with
comparatively smaller predicted habitat area among all the approaches;
this was expected as the expert approach tried to avoid over prediction
while selecting the optimal models (Figure 4a-b). In comparison,
AUCDIFF approaches predicted a comparatively larger
suitable habitat area for a greater number of both fish and odonate
species over the ORTEST approaches (Figure 4a-b). Most
often suitable habitats were over predicted by AUCDIFFapproaches and also for some species by ORTESTapproaches when compared to the area predicted by the EXP approach
(Figure 4a-b). For instance, AUCDIFF_PER and
AUCDIFF_BAL predicted habitat area above 38,000
km2 for four of the odonate species and for one fish
species by AUCDIFF_BAL (Figure 4a-b).
The choice of threshold used to derive binary suitable habitat maps from
the optimal models chosen may explain some of the variation in the area
of habitat predicted. Though we did not assess for the effect of
thresholds used on the habitat areas predicted we present the binary
maps derived using percentile and balance thresholds for four species
with varying occurrence records used for the model building as an
example (Figure 5). In general, we observed the restrictive ‘percentile
threshold’ seemed to restrict the predicted suitable habitat around
occurrence data used for model building, while the less restrictive
‘balance threshold’ seemed to overpredict the habitat for most of the
optimal models selected by sequential approaches (Figure 5).
However, there were statistically significant correlations between the
areas of the optimal models chosen by the EXP approach, with optimal
models chosen by all four sequential approaches (Table S4).
Comparatively, the strongest correlation was with
ORTEST_PER and ORTEST_BAL for fish and
with ORTEST_PER and AUCDIFF_PER for
odonates, though correlation strength differed with very high
correlations for the former and moderate for the latter (Table S4).