We present a novel approach to estimate absence records for species distribution models using rarefaction, a method to estimate the sampling effort. Then, we evaluated possible temporal changes in the suitable habitat of Hypsiboas albopunctatus (Hylidae), a widespread-generalist native frog of the Cerrado ecoregion (South America), using our novel approach to estimate absence records. Finally, we tested our novel approach in another eight frog species. The estimation of presences and absences was based on all frog records of 434 collections worldwide. Presences and absences were classified using 22 environmental predictors. Four classification methods were applied: Maximun Entropy (MaxEnt), Random Forest (RF), Gradient Boosting Machine (GBM), and Support Vector Machine (SVM). Two different approaches were tested to estimate absences: 1) true-absences (our proposed approach) and 2) pseudo-absences (widely used approach). Species Distribution Models (SDM) were built at 2001 and 2013 for H. albopunctatus. We also evaluated our absence approach in another eight frog species following the same methodology described previously for H. albopunctatus.
Results/Conclusions
For H. albopunctatus, RF and SVM generated good model accuracy (AUC > 0.8 and TSS > 0.5) using either true-absences or pseudo-absences. GBM generated good model only using true-absences. The accuracy of MaxEnt models were below acceptance limits for true-absences but good for pseudo-absences (10000 random points by default). RF, GBM and SVM models using true-absences were significantly more accurate than RF, GBM and SVM models using pseudo-absences (AUC: F = 36.26, P < 0.001; TSS: F = 143.1, P < 0.001). Using true-absences, the three spatial models from RF, GBM and SVM classifications indicated that the suitable habitat did not change substantially from 2001 to 2012.
The use of true-absence data produced accurate models for most of the other eight frog species using RF, GBM, and SVM. Only one species did not reach satisfactory AUC (AUC > 0.8) for the three classification methods and neither TSS (TSS > 0.5) for RF and GBM. The rest of the frogs reach satisfactory AUC and TSS for at less two classification methods. We proved that our absence approach allowed confident classification to generate SDM. Our absence approach reduced the uncertainty and bias compared to the most used pseudo-absence approach.