Evaluating the Impact of Climate Change on the Endangered Endemic Species Thuja koraiensis Nakai in Baekdudaegan, South Korea: An Ensemble Modelling Approach

Thuja koraiensis Nakai ( Cupressaceae ) is an endangered conifer species that holds significant ecological importance as an endemic plant in Korea. To facilitate the adaptive management of ecosystems on the Korean peninsula in the face of climate change, the use of species distribution models (SDMs) can be instrumental in supporting climate-adaptive forest restoration programs. In this study


Introduction
Thuja koraiensis Nakai is a short and creeping evergreen shrub that reaches nearly 3 m in height and only grows in northeast China and the high mountains of Korea. (1)Climate change has progressed in recent years and natural populations of T. koraiensis have declined together with the competition with temperate species.Furthermore, the decline in the number of T. koraiensis might be related to inbreeding and inadequate environments for seed germination within the natural habitats situated in restricted areas high atop mountains. (2)Currently, the studies of T. koraiensis are focused on the molecular mechanism of somatic embryogensis, (3) genetic variation, (4) habitat investigation and growth patterns (5) in terms of conservation efforts.The decrease in the habitats of T. koraiensis has emerged as a serious issue in terms of preserving the diverse species on the Korean peninsula. (6)Therefore, the endemic T. koraiensis has been designated as a rare and endangered tree species in Korea and a vulnerable species on the International Union for Conservation of Nature (IUCN) Red List of Threatened Species. (7)pecies distribution models (SDMs) are quantitative models of species-environment relationships typically developed using species location data and those environmental variables considered to affect the species distribution. (8)Commonly, SDMs are used to investigate the environmental suitability to estimate a species' current or future distribution. (9)Furthermore, many studies of future climate change use an ensemble modelling approach in which simulations of future conditions are produced with multiple climate models, rather than just one. (10)As a single algorithm can easily result in the overfitting phenomenon, this problem is eliminated after the ensemble. (11)Hence, ensemble modelling can significantly improve the modelling accuracy and reduce uncertainty compared with modelling using only a single model. (12)However, there is no research on the ensemble modelling of T. koraiensis species in Baekdudaegan (South Korea), which should be predicted with an assessment of their uncertainty, in accordance with the potential impact of climate change.
Therefore, in this study, we use ensemble algorithm SDM methods to evaluate the impact of climate change scenarios under representative concentration pathway (RCP) 4.5 on the endangered native T. koraiensis populations across the South Korean peninsula.Our findings would provide long-term insights into forest conservation and restoration strategies in response to climate change.

Location data for T. koraiensis in South Korea
To acquire data on T. koraiensis in South Korea, we employed databases serviced by Baekdudaegan National Arboretum, Korea Arboreta and Gardens Institute (Gonghwa, South Korea).Four additional field surveys were performed to confirm the current populations of T. koraiensis (Fig. 1).In total, 118 sites where T. koraiensis are present were identified, and maps with binary cell values of presence (1) and absence (0) were produced for each SDM using the threshold value of 0.5 (Fig. 2).

Environmental parameters
Current climate data, including 19 bioclimatic variables in the World Climate data, were used to predict the change in the distribution of potential land suitable for T. koraiensis populations. (13)n particular, bioclimatic variables were considered suitable for investigating the species distributions under current or possible future conditions by applying the SDMs. (14,15)Four (max   with a spatial resolution of 135 km, and its bioclimatic variables were downscaled to a spatial resolution of 30 arcsec using version 1.4 of WorldClim for the actual analysis conducted for our detailed climate forecast in South Korea.

Species distribution ensemble modelling
BIOMOD2 provides an ensemble platform of ten SDM algorithms, and we initially used four of these as ensemble candidates.To model the potential land suitability for T. koraiensis in the future, the ensemble modelling was designed to consist of three different regression models [general additive model (GAM), generalized boosted model (GBM), and general linear model (GLM)] and a machine learning model [random forest (RF), provided by the BIOMOD2 package in R statistical language (Fig. 3). (16)We built individual models using default settings provided by BIOMOD2 version 3.4.1 and ensembled the outcomes of SDM simulations with true skill statistics (TSS) values over 0.8 to reduce the uncertainty of SDMs.This evaluation metric is an  indicator of discrimination capacity and quantifies how well the model distinguishes presence from absence (or presence from background samples, when absence is unavailable).As held-out subsets of the original dataset are used for model evaluation (cross-validation) for partitioning data into subsets.SDM results represent the possibility of each species' occurrence in the form of a continuous distribution, implying the importance of the threshold in objectively determining the presence of a selected species for the integrity of the modelling results and their interpretation.In BIOMOD2, the cross-validation procedure was repeated two times for each of the 10 presence-absence groups. (17)The accuracy of a committee averaging (CV) ensemble methodology was evaluated using outer validation.

Evaluation of the performance of various BIOMOD2 algorithms
The SDM platform and R package BIOMOD2 were used to explore the potential land suitability for T. koraiensis in the future.Also, by evaluating the performance of models, the receiver operating characteristic (ROC) and TSS were confirmed.The distribution habitats of T. koraiensis for both current and future climates were obtained by ensemble forecasting in the BIOMOD2 package by parameterizing TSS > 0.9 for individual models (Table 2).By comparing the TSS values, GBM, GLM, and RF (0.985 ± 0.011 − 0.988 ± 0.011) were found to perform better than the statistical GAM method (0.911 ± 0.109).

Contribution of the four predictor variables used to construct SDMs for T. koraiensis
As shown in Table 3, the importance of variables (environmental parameters) for SDM is presented.Among the four predictor variables under GAM and GLM, bio 5 (0.809 ± 0.066 and 0.997 ± 0.001, respectively) and bio 11 (0.977 ± 0.030 and 0.966 ± 0.054, respectively) contributed the most to the model predictions.In contrast, bio 19 contributed little to the predictive performance of the models among the four predictor variables under the four algorithms (0.086 ± 0.019 − 0.310 ± 0.050).Under GLM, bio 5 (0.997 ± 0.001) and bio 11 (0.966 ± 0.054) contributed the most to the model predictions, followed by bio 7 (0.733 ± 0.131).In summary, the dominant bioclimatic variables shaping the potential distribution of T. koraiensis are bio 5 and bio 11.

Current and future potential distributions of T. koraiensis under RCP 4.5
The predicted distribution habitats of T. koraiensis for both current and future climates were revealed by the ensemble prediction (Fig. 4).The natural population of T. koraiensis currently spreads throughout Baekdu-degan including Gariwangsan, Gyebangsan, Jangsan, Taebaeksan, Hwahangsan, and Seoraksan.The predicted ensemble-averaged model showed the area loss of currently suitable habitats for T. koraiensis, yet the species is predicted to still exist in the future under RCP 4.5.
Under the RF models, the potential habitat distribution of T. koraiensis would be conserved under RCP 4.5 in the 2030s, 2050s, and 2070s.In particular, in the 2070s, an increase in the number of natural habitats of T. koraiensis is predicted.Similar to RF models, GLM models showed expanded natural populations of T. koraiensis throughout the Korean peninsula.Specifically, the areas near Mt.Halla on Jeju Island and Jirisan were identified as new potential habitats in the 2030s and 2050s.Yet, these predicted potential habitats of T. koraiensis were excluded in the 2070s owing to habitat loss across the Korean peninsula when considering the climatic conditions under RCP 4.5.The prediction results obtained with GAM showed a trend different from those obtained with RF and GLM.GAM predicted that the potential habitat for T. koraiensis will greatly expand in the 2030s owing to climate change.However, this expansion of natural habitats of T. koraiensis would decrease in the 2050s and 2070s.Similar to GLM, predicted potential habitats such as Mt.Halla on Jeju Island and Jirisan would be excluded in the 2070s.

Discussion
Climate has been altering rapidly in South Korea.Specifically, it is predicted that by 2070, the average temperature will increase by 1.3 to 5.2 ℃, and the average temperature will increase by up to 29.6%. (18)Climate change is an important factor that regulates the growth and survival of trees. (19)SDMs are widely used to support the climate-adaptive program of threatened species as they suggest potential translocation sites. (9)lobal warming is altering the natural habitat environments, threatening the natural populations of T. koraiensis in South Korea.In this study, we used the multiple SDMs for the ensemble average method and predicted that T. Koraiensis is vulnerable to future climate change.Several algorithms in this study performed well with our data, showing that the results of using the four methods converged successfully with stable TSS results.TSS provides a threshold-dependent measure of accuracy and is readily applied for presence-absence predictions. (20)limate projections using three different models (RF, GLM, and GAM) were used to assess the potential habitat of T. Koraiensis under RCP 4.5.On the basis of the results of the screening of the SDMs, RF is selected to construct a high-precision ensemble model as it showed the highest TSS (0.988 ± 0.011).The common finding obtained using the three models was that climate change in the Korean peninsula under RCP 4.5 has an impact on the natural habitat of T. koraiensis, as shown by the area loss of the potential population.Above all things, the species is predicted to still exist in the 2070s under RCP 4.5.These findings will provide meaningful insights for the forest conservation and restoration strategies in response to climate change.
The conservation of endangered species needs to take full consideration of the potential habitat of the species and plant saplings in the right locations. (21)By analyzing the potential habitat of T. koraiensis in ensemble models, we found that there is still a significant amount of land with good potential for restoration.In this study, only climate factors were considered in the analysis, but further research is needed to include topographic factors such as altitude and soil information, as well as the various environmental factors that are suitable for the growth environment characteristics of T. koraiensis.Additionally, in order to conserve the natural habitat of T. koraiensis on the Korean peninsula, it is necessary to establish spatial information data for regions with similar suitable environmental conditions for T. koraiensis as potential distribution areas.Using the established database, we must initiate efficient management strategies for the protection and nurturing of T. koraiensis in response to climate change.

Conclusions
Climate change will affect ecologically sensitive areas, so it is crucial to manage its impact on various species.T. koraiensis is an endangered and ecologically important conifer endemic to South Korea.SDMs can support climate-adaptive forest restoration programs by predicting the effect of climate change on the future dispersal and distribution of targeted species.In this study, we used multiple SDMs in the ensemble average method for predicting the changes in the endangered native T. koraiensis populations in response to climate change.The predicted ensemble-averaged model showed the area loss of currently suitable habitats of T. koraiensis, yet the species was predicted to still exist in the 2070s under RCP 4.5.Furthermore, several algorithms in the BiOMOD2 ensemble performed well with our data, showing that the four methods converged successfully, presenting stable TSS results.Our study suggests that common results from different SDMs can be used as the basis for developing conservation strategies to protect the native rare T. koraiensis in South Korea.These findings would provide a long-term perspective for preparing protective strategies of vulnerable species against climate change.

Fig. 3 .
Fig. 3. (Color online) Schematic representation of the ensemble species distribution modelling approach.
temperature of warmest month: bio 5, temperature annual range: bio 7, mean temperature of the coldest quarter: bio 11, prediction of the coldest quarter: bio 19) of the 19 BioClim variables were considered when applying the SDMs (Table1).For future climate data, we referred to the HadGEM2-AO climate model constructed by the National Institute of Meteorological Science for the publication of the Intergovernmental Panel on Climate Change (IPCC) Assessment Report 5 (AR5) for the 2050s (2041-2060) and 2070s (2061-2080) under the RCP 4.5 emission scenario (RCP 4.5).The HadGEM2-AO climate model is a general circulation model (GCM)

Table 1
Bioclimatic variables used for species distribution models to predict the potential future distribution of targeted species.

Table 2
Mean and standard deviation of ROC and TSS for the models.
Each value is expressed as mean ± SD.