3.
Results
In total, for each of the random and background weighting schemes we
fitted 128 preliminary models based on the 10 cross-validated folds of
the training occurrence points. Since the GLM is inherently a simple
algorithm, for this method only one set of parameters was trained.
Although for both traning and test dataset the AUC and TSS of this model
in BkWt scheme decreased in comparison to random scheme (Table 1 & 2),
it successfully classified all the test data, i.e., sensitivity = 1, by
using the BkWt scheme. Totally, GLM obtained AUC of 0.92 and 0.89 and
TSS of 0.66 and 0.65 for random and BkWt schemes, respectively (Table
1). From the multiple combination of the GBM parameters, for the random
background selection a model characterized by shrinkage = 0.01,
interaction depth = 5, and ntrees = 1800 showed highest ROC value (ROC =
1, sensitivity = 1, and specificity = 0.85). For BkWt, the fine-tuned
GBM model (ROC = 1, sensitivity = 1, and specificity = 0.64) had
shrinkage = 0.01, interaction depth = 5, and ntrees = 2000. Although for
both training and test datasets the AUC and TSS of this model were
almost equal (AUC = 0.976 and 0.971 and TSS = 0.59 and 0.58 for random
and BkWt schemes, respectively), it lost the ability to truly predict
presence points, i.e., sensitivity, compared to othe models (Table 2).
For the RF model, the fine-tuned model of both random and BkWt schemes
was charachterized by ntrees = 1000, and nodesize = 1. However, the mtry
was 2 and 3 for random and Bkwt schemes, respectively. Similar to the
GBM, RF obtained almost equal AUC and TSS scores for both training and
test datasets (AUC = 0.97 and 0.96 and TSS = 0.45 and 0.58 for random
and BkWt schemes, respectively), but the sensitivity of this model was
low (Table 2). For the MaxEnt model the best-fitted model with the
highest AICw of the ENMeval analysis obtained rm 0.5 and
1.5, and fc LQ and LQHP for random and BkWt schemes, respectively. In
the final habitat suitability maps of the MaxEnt model, the test data
had AUC 0.93 and 0.95 and TSS 0.66 and 0.80 for random and BkWt schemes,
respectively (Table 2). Overall, the highest TSS score of the test data
was obtained in the MaxEnt model that was fitted based on the BkWt
background selection scheme (Table 2).
The predicted suitability maps are shown in Fig. 2. We found a good
consistency between the patterns of occurrence points and suitable
habitats. Comparing the spatial pattern of suitable habitats in random
and background weighting methods showed that all models represented
different results except for MaxEnt model in which comparable results
were obtained (Fig. 2). Accordingly, we calculated the correlation
coefficient between the two background selection schemes of the four SDM
methods revealing that the highest correlation was obtained for MaxEnt
model (r = 0.85), followed by GLM (r = 0.61), GBM
(r = 0.45), and RF (r = 0.42). These findings were also
confirmed by sensitivity and specificity graphs (Fig. 3). We found that
while the capacity of GBM and RF to predict the training and test
background points (i.e. models’ specificity) was maintained excellent
even at higher thresholds, their capability to predict presence data
(i.e. models’ sensitivity) were reduced at lower thresholds. On the
contrary, GLM and MaxEnt models showed good performances to predict
presence data, but lost their capacity to classify background data at
lower thresholds especially in the BkWt scheme (Fig. 3). The comparison
of the response curves of the variables between the two bakhground
selection schemes indicated identical pattern, however, for GBM and RF
the response curves of the BkWt scheme were more rugged compared to the
smoother variation in the random background selection scheme (Fig. 4).