5. Conclusion
Spatial bias of the input data is one of the main sources of uncertainty
in the species distribution modeling approaches. This issue is
particularly important for scarce species with geographically imbalanced
biased data of their distribution ranges. Despite the great emphasis on
the importance of model tuning and input data manipulation in improving
SDMs, the performance of different models in using such an approach has
not received much attention. In this research we evaluated the
performance of four commonly-used SDMs to predict imbalanced biased
occurrence points based on two methods of background data selection
including random and background weighting. Our result reveals that
different models produced dissimilar results for two background
selection schemes. Complex GBM and RF models, due to their interpolative
conception, showed inefficiency in predicting test points, especially
for the background weighting mode. The GLM over-predicted presence areas
due to its extrapolative nature. In spite of being a machine learning
method, MaxEnt show a comparable performance in predicting test points
in two background selection schemes. The results of the present study
emphasize the proficiency of MaxEnt model in generating reproducible
comparisons particularly when the input data is being completed.