Figure legends
Figure 1. Schematic diagram of selecting optimal models. Steps with red
arrows are unique to AUCDIFF approaches while all other
steps are common to all the four sequential approaches. Green boxes with
“Single model selected” are the optimal models for either
ORTEST or AUCDIFF approaches. Step 2a is
followed then one will derive optimal models for AUCDIFFafter Step 2, otherwise optimal models derived will be for
ORTEST approaches. While purple box with “Single model
selected” is the optimal model for only AUCDIFFapproaches. Step 5 is the ultimate step wherein models with lower
feature class is chosen as the optimal model when multiple models have
same numbers of parameters and average absolute value.
Figure 2. Regularization multiplier and feature class (RM-FC)
combinations for (a) fish and (b) odonate of the optimal models chosen
by five model selection approaches.
Figure 3. Summary of different features of the optimal models selected
through five optimal model selection approaches for the fish and odonate
species of Bhutan. 10 percentile training presence test omission (1) and
balance training omission, predicted area and threshold value test
omission (2) for the (a) fish and (b) odonate; AUC difference for the
(c) fish and (d) odonate; average number of parameters for the (e) fish
and (f) odonate; test AUC for the (g) fish and (h) odonate species of
Bhutan. Expert approach based on ecological plausibility of binary
suitable/unsuitable model (EXP), sequential approaches using 10
percentile training presence test omission and test AUC
(ORTEST_PER), 10 percentile training presence test
omission, AUC difference and test AUC (AUCDIFF_PER),
balance training omission, predicted area and threshold value test
omission and test AUC (ORTEST_BAL) and balance training
omission, predicted area and threshold value test omission, AUC
difference and test AUC (AUCDIFF_BAL).
Figure 4. Area of the predicted
habitats of the optimal models selected through the five model selection
approaches for the (a) fish and (b) odonate species of Bhutan.
Figure 5. Examples of binary suitable habitat maps derived using 10
percentile training presence Cloglog threshold (Left panel) and balance
training omission, predicted area and threshold value Cloglog threshold
(Right panel) among the optimal models chosen by the Expert and the four
sequential optimal model selection approaches. (a) Cyprinion
semiplotum (5 occurrence; EXP, ORTEST_PER,
ORTEST_BAL, AUCDIFF_PER and
AUCDIFF_BAL), (b) Neolissochilus hexagonolepis(12 occurrence; EXP, ORTEST_PER); (c)Neolissochilus hexagonolepis (12 occurrence;
ORTEST_BAL, AUCDIFF_PER and
AUCDIFF_BAL); (d) Aristocypha quadrimaculata (8
occurrence; EXP, ORTEST_PER,
ORTEST_BAL, AUCDIFF_PER and
AUCDIFF_BAL); (e) Diplacodes trivialis (21
occurrence; EXP); (f) Diplacodes trivialis (21 occurrence;
ORTEST_PER); (g) Diplacodes trivialis (21
occurrence; ORTEST_BAL, AUCDIFF_PERand AUCDIFF_BAL).