Modeling the current and the future distribution of Anatolian chestnut (Castanea sativa mill.)

With the increase in number of statistical methods and geographic information systems tools, the use of "species distribution models" in ecology has gained momentum. Today, species distribution models are widely used in different fields such as biogeography, conservation biology, climate change research, ecological restoration, habitat or species management. Modeling the current and future distribution of Anatolian Chestnut (Castanea sativa Mill.), provided us important information about the habitat of this species. By different modelling techniques, taking the climate change into account, maps and supplementary materials would be also created to support current and future afforestation/silviculture plans, ecological restoration, intervention methods against invasive species and detection of nature conservation areas. The distribution of sweet chestnut in Turkey has been critically disturbed, particularly in the Marmara and Aegean regions, due to the negative effects of "Chestnut Blight" (Cryphonectria parasitica), "ink disease" (Phytophthora cambivora), and more recently, the "Asian Chestnut Gall Wasp" (Dryocosmus kuriphilus) (Vannini & Vettraino, 2011; Dal Maso et. al, 2017; Bombi et. al., 2019). These pathogens have caused significant tree deaths and crop losses in Türkiye (Bucak, 2006; OGM, 2012). Species Distribution Models (SDMs) can be used to create risk maps of these pathogens, helping to locate damaged chestnut forests for ecological restoration and prevent total degradation (Jiménez-Valverde et. al., 2011; Cunniffe et. al., 2016). According to FAO reports in 2012, global chestnut production averages 2 million tons per year. China leads with 1.3 million tons (84%), followed by Turkey with 60,000 tons (3%), and then Italy, South Korea, Bolivia, and Greece. The 2013-2017 "chestnut action plan" by the General Directorate of Forestry states that Türkiye has 28,804 hectares of pure chestnut, including 22,011 productive and 6,793 degraded stands. In addition to pure stands, chestnuts are mixed with many different species of broad-leaved and coniferous; these mixtures are divided into areas where chestnut is the primary or secondary species. Sweet Chestnut mixtures cover 233,241 hectares, with about 20,000 hectares being degraded stands. In this study, chestnut distribution was modeled at two different spatial resolutions: national (1km) and regional (90m) scales. The current national-scale models were created using two different datasets (WorldClim 2.1; 1970-2010 and CHELSA 2.1; 1981-2010). For both scales, three different algorithms (Random Forest, Gradient Boosting Machine, and Generalized Linear Models), ten different pseudo-absence datasets, and the spatial cross-validation method (blockCV; Valavi et al., 2019) were used for model calibration and validation. Future projections for the national-scale models (2021-2040 and 2041-2060 periods for WC2.1; 2011-2040 and 2041-2070 for CH2.1) were created using global circulation scenarios from the CMIP6* (Eyring et. al., 2016; O'Neil et. al., 2016) research groups of the GFDL, IPSL, and MPI institutes. Changes between the two different "current habitat suitability" maps and six different "future habitat predictions" maps at the national scale were calculated using the SRC (Species Range Change) method and presented in tabular form within the study. The results indicate that the potential distribution maps obtained from the CHELSA 2.1 dataset provide more optimistic results compared to the WorldClim 2.1 dataset. The regional scale models are based on the hierarchical modeling approach proposed by Mateo et al. (2019). In this phase, the resolutions of the bioclimatic layers obtained from WC2.1 used in the hierarchical modeling were sampled at the same resolution (90m) as the morphological variables by using Geographically Weighted Regression (CAR; Brunsdon et al., 1996) method. The evaluation of the models applied at both scales resulted in Receiver Operating Characteristic (ROC) scores ranging between 95-99% and True Skill Statistics (TSS) scores ranging between 70-85% for each algorithm, indicating satisfactory model accuracies were achieved. The final maps, including both algorithm-specific and ensemble models, were presented in the thesis.