Climate change and bird extinctions in the Amazon

correlation test) to identify potential collinearity between these variables (S1 Table) using the R software v3.4.0 [59]. When the correlation between variables was high (r> | 0.8 |), we did not use one of them to generate the distribution model. Thus, we selected Average annual temper- ature (Bio1), Isothermality (Bio3), Temperature seasonality (Bio4), Annual precipitation (Bio12), and Precipitation of the wettest month (Bio13). To future climate scenarios (2050), we used the same variables selected for the current scenario and estimated models for 11 Atmo- spheric-Ocean General Circulation Models (BCC-CSM1-1, CCSM4, GISS-E2-R, HadGE- M2-AO, HadGEM2-ES, IPSL-CM5A-LR, MIROC5, MRI-CGCM3, MIROC-ESM-CHEM, MIROC-ESM, NorESM1-M), which are all available at the Worldclim database. We consid- ered both the optimistic (rcp45) and the pessimistic (rcp85) estimates of greenhouse gas emis- sions. The rcp45 depicts the mitigation scenario, while the rcp85 is the baseline business-as- usual scenario without additional efforts to restrain gas emissions [60]. Modeling and model evaluation Potential species distribution modeling was performed using three different algorithms to obtain more reliable predictions of the species distribution [61]. Maximum Entropy (MaxEnt) [62] is based on a presence-background approach that evaluates the relationship between the actual species occurrence and the entire study area [42]. On the other hand, Support Vector Machine (SVM) [63] and Random Forest (RDF) [64] algorithms are presence-absence models

that use occurrence records and compare them with absence data. For algorithms that require pseudo-absences, we used the random space allocation method [65]. The number of pseudo- absences was equal to the number of presences for each species. In order to produce more accurate models, we use the bootstrap method [66] to partition data into 70% training and 30% testing sets and repeated this procedure five times. Species Distribution Modeling analysis included in the package ENMTML [67] scripts available on GitHub (github.com/andrefaa/ ENMTML). All models were generated using the R software. We evaluated models using the Sorensen index [68, 69], which ranges from 0 to 1. It is a threshold-dependent method where values closer to 1 indicate greater overall model accuracy; i.e., it has high precision and good performance [69]. Subsequently, we combined all model predictions for a species to obtain its final model (full ensemble) for both the present and future scenarios. We obtained the final ensemble through the weighted average of all generated models whose values were higher than the mean Sorensen values. Then, we applied a spatial restriction method based on species dispersal limitation [70, 71]. This approach—called here MSDM—relates the likelihood of a species’ occurrence at a given site to its distance from other localities where the species has been documented. In this method, we first created a layer of the summed distance from each cell to all occurrences that were later incorporated into the model fitting [70, 72]. This procedure is available in the script

used in all of our modeling process (github.com/andrefaa/ENMTML) [67] (Fig 2). Overlap of species distribution models and deforestation models Knowing that these are forest-dependent species, we overlapped the SDMs with forest rem- nants for both present and future scenarios (Fig 2). We obtained information on currently for- ested areas from the Brazilian Amazon Rainforest Satellite Monitoring Program (PRODES; http://www.obt.inpe.br/OBT/assuntos/programas/amazonia/prodes) from the National Insti- tute for Space Research of Brazil (INPE). Data on future forest cover have been extracted from the SimAmazonia database (http://csr.ufmg.br/simamazonia/, [26]), which provides two types of deforestation scenarios: business-as-usual (BAU) and governance (GOV). The BAU sce- nario assumes that novel protected areas will not be created, a low compliance rate with PLOS ONE Bird extinctions in the Amazon PLOS ONE | https://doi.org/10.1371/journal.pone.0236103 July 17, 2020 5 / 18