2.4 Potential geographic
distribution
ENM or species distribution models (SDM) can predict potential spatial
distributions of species. Generalized additive models (GAM), random
forests (RF), boosted regression trees (BRT, or named GBM), maximum
entropy (Maxent) are widely used models (Guisan et al., 2014). Although
there are many model options, no single optimal metric is widely
applicable in this field (Qiao et al., 2015). The consensus algorithm
can balance the performance of multiple models (Marmion et al., 2009),
but results are ambiguous compared to a single model (Breiner et al.,
2015; Zhu & Peterson, 2017). In order to obtain an optimal result
predicting potential distribution, three individual models (RF, GBM,
Maxent) and an ensemble model were implemented in the biomod2 package
(Thuiller et al., 2009). Seventy percent of occurrence data was used for
model training and 30% for model testing. We selected the partial
receiver operating characteristic (PROC) as the model evaluation
criteria; in contrast with the AUC method, PROC eliminates the
misleading effects of absent data on the results and emphasizes the
crucial role of the omission rate to prediction performance (Peterson,
2006). An AUC ratio of 1 implies that the niche model is no better than
a random prediction, and a larger AUC ratio indicates better
discrimination in the partial ROC approach (Peterson et al., 2008). In
addition, given criticism of the complexity and transferability of
Maxent default settings, we adjusted regularization multiplier (RM)
values and feature combination (FC) settings in the ENM eval package to
optimize parameters and determined the delta AICc minimum and average
AUCtest maximum values to generate Maxent (see Figure S3) (Muscarella et
al., 2014).
A more complete niche assessment can be obtained using all species
distribution data (Broennimann & Guisan, 2008), so we used total
occurrence data to model and analyze the Asian openbill potential
distribution. First, we generated a calibration model using the present
data within calibration range, then the potential distribution was
predicted to a new projected range. Due to the potential extrapolation
uncertainty of the model in the transfer, we still used ExDet to
determine the novel environmental parameter (Type 1 novelty)(see Figure
S2.3) (Mesgaran et al., 2014).
To determine whether the probability of two new distribution areas was
higher than other areas where the Asian openbill did not yet occur, we
randomly created occurrence sites outside of the existing distribution
area within the projected extent. A probability value of 0-1 was
generated in the optimal model with the highest PROC value to compare
the occurrence probability of species in the native area, new
distribution areas, and absence area.