Past, present and future distributions of Oriental beech (Fagus orientalis) under climate change projections

parameters (see S1 Table). The WorldClim dataset offers global climate model (GCM) simula- tions for the past (LGM and MH) and future (2050 and 2070) climatic conditions. We used MIROC-ESM and CCSM4 global climate simulations since they were the only two GCMs that provided data for all focal periods [40,41]. In addition, for the future simulations, climate sce- narios based on IPCC reports were used [2]. According to these scenarios, in 2100, the expected CO2 levels will be 450 ppm, 650 ppm, and 1350 ppm in optimistic, moderate, and pessimistic scenarios, respectively. In addition, the expected increase in mean annual tempera- ture will be 0.2–1.8˚C, 1.0–2.6˚C, and 2.6–4.8˚C, in optimistic, moderate, and pessimistic sce- narios, respectively [2]. To avoid issues arising from predictor collinearity, we retained only those bioclimatic variables with low correlation (|r| < 0.85) (Table 1, [42,43]). As a result, we used a subset consisting of nine bioclimatic variables (Table 2). In addition, the particular ecol- ogy of beech forests was taken into account when reducing the collinearity within the environ- mental parameters. According to Salamon-Albert et al. [44], temperature variables (annual mean, warmest, and coldest seasonal values) determine the distribution of beech forests. For Oriental beech forests at the latitudes covering our study region both temperature and precipi- tation factors play important roles. In addition, accounting for both temperature and precipi- tation in the model structure is important to consider water and moisture availability [45].

Model structure and analysis SDMs can describe or predict the probability of the presence or absence of a species across environmental gradients or in a specified geographical area [46]. We used the R-package ‘bio- mod2’ version 3.3–7.1 [47] to create potential-distribution maps of F. orientalis. To evaluate model performance, we implemented five different algorithms commonly used for SDMs pro- vided with ‘biomod2’: Generalized linear model (GLM), general additive model (GAM), ran- dom forest (RF), BIOCLIM, and maximum entropy (MaxEnt). MaxEnt uses presence-only data. In contrast, GLM, GAM, and RF require both presence and absence data, and BIOCLIM uses pseudoabsence data [47], thus, we introduced 20,796 random background points to increase precision, as these add uniformity to the model while still subjected to climatic con- straints [45]. Default settings were used for data formatting; 70% of the input data was used as a training sample. The related code is provided in the supplementary (S1 File). Past and future climates were projected through the trained model, which is based on pres- ent conditions with observed distribution data. To evaluate and compare the model perfor- mances we used the Receiver Operating Characteristic Curve (Area Under the Curve, AUC) which takes values between 0 and 1 (AUC > 0.5 indicating that the model performed better than random), higher AUC value meaning better model performance [48]. Model outputs are given as the species’ presence probability in all algorithms.

To display and evaluate possible distributions we used both lowest predicted value [49] and sensitivity–specificity equality [50] methods to set a threshold for the minimum predicted value of observed occurrences, which was calculated as 0.585. We aimed at zero omission by setting this threshold and tried to balance sensitivity and specificity for increased accuracy. Fossil pollen data To compare and test model performances for past projections we compiled published palyno- logical records of Fagus comprising LGM and MH for the range of F. orientalis. Although Fagus pollen had low values in several pollen diagrams pointing to extra-local origin of the mother plant (e.g. [51]), we accepted such records as evidence for presence of oriental beech in the region during target periods (LGM, MH). PLOS ONE Species distribution modeling of Oriental beech (Fagus orientalis) PLOS ONE | https://doi.org/10.1371/journal.pone.0242280 November 17, 2020 4 / 19