Pure, shared, and coupling effects of climate change and sea level rise on the future distribution of Spartina alterniflora along the Chinese coast

Abstract Aim Global change seriously threatens the salt marsh ecosystem, while it remains unclear how S. will respond to climate change and sea level rise. Here, we investigated interactions among variables and identified the impacts of climate change, sea level rise, and their interactions on the distribution of Spartina alterniflora. Location Northern Chinese coast and Southern Chinese coast. Taxon Spartina alterniflora Loisel. Methods With global sensitivity analysis, we determined interactions among variables and their relative importance to the distribution of S. alterniflora. Integrating the Venn's four‐set diagram, we built ecological niche models under current and three future scenarios to identify pure, shared, and coupling effects of climate change and sea level rise on the distribution of S. alterniflora. Results Mean diurnal range (Bio02) and Elevation were the two most critical variables controlling the distribution of S. alterniflora on the Chinese coast, and interactions among variables of the northern coast were much greater than that of the southern coast. Habitats change was mainly caused by pure effects of climate change, except habitats reduction on the southern coast. Pure effects of sea level rise were low, but it can influence habitats change through shared and coupling effects from complex interactions with climate change. Interactions of climate change and sea level rise can drive habitats change, and the changed habitats caused by shared and coupling effects were mainly distributed the areas near the landward side. Main conclusions Our research suggests paying attention to interactions among variables when calculating the relative importance of explanatory variables. Identifying pure, shared, and coupling effects of climate change and sea level rise for the distribution of S. alterniflora will provide scientific references for assessing the risk of similar coastal species.


| INTRODUC TI ON
It is now well established that the earth's climate is warming (Cazenave & Cozannet, 2014;Priest et al., 2010) and that the rate of sea level rise is accelerating, with a projection of global sea level rise of 0.75-1.90 m by next century (Team et al., 2014). Although previous studies have reported the impacts of projected sea level rise on salt marsh plant communities (Allen & Lendemer, 2016;Donnelly & Bertness, 2001;Kirwan et al., 2010;Valle et al., 2014), few investigators have examined the interactive effects of sea level rise and climate change for the salt marsh plants (Garner et al., 2015;Kirwan & Mudd, 2012). Their interactions and compounding effects may lead to reduced salt marsh sustainability (Charles & Dukes, 2009;Cherry, Mcknee, & Grace, 2009) and influenced the ability of salt marsh plant to survive (Kirwan et al., 2013). Kirwan (2012) also found that plants responses differed depending upon the elevation of the marsh relative to sea level under interactive effects. Recent studies showed that interactions may further change the composition of species assemblages and making important ecological processes at salt marshes uncertain in the future (Garner et al., 2015;Hanson et al., 2016). So far, there has been little discussion about habitats change caused by the pure effects of climate change, sea level rise, and the shared and coupling effects from their interactions. Quantifying these effects will help us better understand the effects of climate change and sea level rise on the coastal ecosystem, and further accurately assess the risk caused by them.
Coastal ecosystems are expected to be exposed to the increased risk of experiencing adverse consequences related to climate change and exacerbated by rising sea level (Nicholls et al.,2007;Valle et al., 2014), while it is still unclear how coastal ecosystems respond to them. It is possible that in coastal ecosystems, native species will decline due to their poor adaptability to these threats (Mendoza-Gonzalez et al., 2013). However, invasive exotic species may take the opportunity to expand their habitats and stabilize their colonial status. So, understanding how invasive coastal species respond to climate change and sea level rise is becoming an urgent challenge (Brierley & Kingsford, 2009;Hoegh-Guldberg & Bruno, 2010).
Spartina alterniflora Loisel, native to the Atlantic and Gulf coasts of North America (Wang et al., 2006), is a highly invasive species widely distributed along the Chinese coast (Yang et al., 2008). A large number of published studies have already revealed the physiological characteristics and expansion mechanisms of S. alterniflora based on laboratory work (Deng et al., 2006;Hu et al., 2015;Shi et al., 2007;Wang et al., 2015;Gu & Zhang, 2009;Zhao et al., 2007;. These studies demonstrated that the expansion of S. alterniflora is influenced by elevation, climate, soil salinity, inundation duration, pH, and many other variables could influence, but previous studies only focus on the single-variable and ignore the interactive effects of multiple variables (Braun, Schindler, & Rihm, 2017;Daniel, Hubert, GertJan, & Wim, 2009). Although studies have recognized the importance of interactions, fewer researches have systematically identified interactions among variables (Liu et al., 2018). Moreover, as a salt marsh plant, the distribution of S. alterniflora is peculiarly prone to the impacts of coastal change. S. alterniflora is correlated with variations in sea level, and its productivity peaks at intermediate elevations within the intertidal zone (Kirwan et al., 2013). However, no research has surveyed the response of S. alterniflora to climate change, sea level rise, and their interactions. Through the maximum entropy model, we built ecological niche models to explore the response of S. alterniflora to climate change, sea level rise, and their interactions under three future scenarios (only considering climate change, only considering sea level rise and both considering climate change and sea level rise) on the northern and southern Chinese.
Specifically, the following issues will be addressed:

| Study area
The geographical extent of the study area was obtained by a 50-km inland buffer of the shoreline of China including 14 Chinese provinces. According to different coast types and colonization characteristics of S. alterniflora (Gao et al., 2014), the study area was divided into two regions: northern Chinese coast and southern Chinese coast ( Figure 1).

| Data sources
Most of presence records of S. alterniflora on Chinese coast were obtained from published studies Xie & Gao, 2009;Zhang et al., 2010;Zhang et al., 2008;Zhao et al., 2015;Zheng et al., 2018). The others were obtained from field sampling, and the Global Biodiversity Information Facility (available at http://data.gbif.org/).
The presence records were resampled in ArcGIS 10.2 to ensure that there is only one observation within the 1° by 1° cell to avoid spatial autocorrelation and reduce sampling bias (Merckx et al., 2011),  et al., 2005). Bioclimatic data include two groups, one of which is 19 bioclimatic under current conditions and the other is 19 bioclimatic variables of future climatic conditions (RCP 8.5: A scenario of comparatively high greenhouse gas emissions, Riahi et al., 2011).
Considering the collinearity among bioclimatic variables may lead to overfitting, we used the following measures to reduce the number of variables. Firstly, using Spearman rank correlation coefficients, we eliminated bioclimatic variables with the highest and most significant correlation coefficients (|r| > 0.8 and p < 0.001) (Supporting information Figure S1). Then, Boruta, a wrapper built around the random forest classification algorithm implemented in the R, was used to select variables according to the relative importance of bioclimatic variables (Supporting information Figure S2) (Kursa, Jankowski, & Rudnicki, 2010

| Ecological Niche Modeling based on MaxEnt
MaxEnt, one of the most popular machine algorithms, is designed for modeling the geographical distribution of species from the n-dimensional environmental variables spaces with presence-only data (Phillips et al., 2006). A 10-fold cross-validation procedure, which is preferable to penalty functions for assessing model generality, was implemented to replicate model runs and data partitions (Merow et al., 2013). It holds out 10% of the data as a testing set at each of 10 iterations, training the model on the remaining 90% of the data in each iteration. Other specified parameters and their setting are maximum number of background points = 10,000, Maximum iterations = 1,000, Convergence threshold = 0.00001, prevalence = 0.5.

| Global sensitivity analysis based on FAST method
Sensitivity analysis (SA) is the study of how the uncertainty in the output of a model can be apportioned to different sources of uncertainty in the model input (Saltelli & Homma, 1992) and is usually divided into local sensitivity analysis and global sensitivity analysis. Compared to local sensitivity analysis, the range of variation of the parameters of the global sensitivity analysis can be expanded to the entire domain and interactions between parameters can be considered (Haaker & Verheijen, 2004). The Fourier amplitude sensitivity test (FAST) F I G U R E 1 Study area. (a) the northern Chinese coast (b) the southern Chinese coast method, based on performing numerical calculations to obtain the expected value, is more efficient to calculate sensitivities than other variance-based global sensitivity analysis methods (Dan, 2010;Mcrae et al., 1982;Saltelli & Bolado, 1998). The FAST method give first-order sensitivity indices (S F ) and total sensitivity indices (S T ) using the terms in the Fourier decomposition of the model output. The S F measures the main effect contribution of each variable to the total output variance, and the S T accounts for the total contribution including main effects and interactions effects (Vanuytrecht et al., 2014). The difference between S T and S F which can assess the impacts of interactions among variables (Nossent, Elsen, & Bauwens, 2011).
In order to identify interactions among variables and their relative importance for the distribution of S. alterniflora on different regions, we performed the FAST sensitivity analysis with Simlab software version 2.2 (Joint Research Centre of the European Commission, 2011). The probability distribution functions were generated for all variables and significance levels were set at the 1% level. All fitted results passed the chi-square test (p < 0.005) (Supporting information Table S1 and S2). The method of Fast (Saltelli et al., 1999) and N is the total number of parameter sets and model executions. A total of 49,935 input parameter sets were generated using probability distribution functions on different regions with Simlab software version 2.2. These parameter sets were run in MaxEnt and then used for global sensitivity analysis in Simlab.

| Model tuning and evaluation
Feature types combination (FC) and regularization multiplier (RM) are two important parameters that affect model complexity (Merow et al., 2013;Muscarella et al., 2015). ENMeval, an R package, was proved to be useful for tuning these two parameters (RM and FC) (Muscarella et al., 2015). Thus, corrected Akaike information criteria (AICc) value was used to estimate the model complexity in MaxEnt (Dan & Seifert, 2011).The smallest AICc was chosen for model simulation and was thought to can reduce model complexity relative to the default model.

The threshold-independent and threshold-dependent measures
were used to evaluate model performance. Area under the curve (AUC) metric, a typical threshold-independent measure, was utilized as a measure of model accuracy ( Lobo et al., 2008). Values of AUC generally range from 0.5 (equivalent to that due to chance) to 1.0 (perfect performance). Values > 0.9 are considered good, 0.7-0.9 are moderate, and <0.7 are poor (Fielding & Bell, 1997). The true skill statistic (TSS), a commonly used threshold-dependent measure of model accuracy (Allouche, Tsoar, & Kadmon, 2006), is calculated as sensitivity + specificity −1. Values > 0.6 are considered good, 0.2-0.6 are fair to moderate, and < 0.2 are poor (Allouche et al., 2006).

| Identifying pure, shared, and coupling effects of climate change and sea level rise on species distribution
To explore the impacts of climate change, sea level rise, and their interactions on the distribution of S. alterniflora, we designed three (1) The Venn's four-set diagram. Based on the current habitats and future habitats, we could find the changed habitats (H changed = b+c + e+n + i+j + h+k) and unchanged habitats (H unchanged = a+d + f+g). For the changed habitats, we divided it into the increased habitats (H + changed = e + c + n + b) and decreased habitats ( For the increased and decreased habitats, they could be spilt into 4 parts. We define e (H + For the changed habitats of S. alterniflora, it was spilt into eight parts in Equation (2). For c and j, they belonged to the habitats when just considering sea level rise, so they were assumed as increased and decreased habitats caused by pure effects of sea level rise (H p_slr  Finally, according to Equation (3-12), we quantified pure, shared, and coupling effects of climate changes and sea level rise and identified their spatial distribution using spatial analysis tool in ArcGIS 10.2. (2) The results of model evaluation showed that the AUC values were all greater than 0.9 and the values of TSS (three threshold rules such as the maximum training sensitivity plus specificity cloglog threshold, 10% training presence cloglog threshold, and equal training sensitivity and specificity cloglog threshold) were greater than 0.7 in Figure 4. So all the models performed well.

| Interactions among variables and their relative importance for the distribution of S. alterniflora on different regions
As shown in Moreover, the sum of S F is 0.9289 and the sum of S T was 1.3310.
It indicated that there were still interactions among variables, but much weaker than that on the northern Chinese coast. Although the S F and S T of Elevation were much higher than Bio02, the difference was as great as that of Bio02, which indicated that both of them had distinguished interactions, especially for Bio02.

| Pure, shared, and coupling effects of climate change and sea level rise on the distribution of S. alterniflora
As shown in Table3

| Spatial distribution of changed habitat of S. alterniflora caused by pure, shared, and coupling effects of climate change and sea level rise on different regions
Along the northern Chinese coast (Figure 5a), the decreased habitats caused by pure effects of climate change mainly distributed on the Bohai bay, Laizhou bay, and Yangtze River Estuary, while the increased habitats mainly distributed on the Shandong Peninsula, Liaodong Peninsula, and northern Jiangsu Province.
The changed habitats caused by shared and coupling effects were small and it mainly distributed in the coastal zone of Jiangsu province. On the southern Chinese coast (Figure 5b), the increased habitats caused by pure effects of climate change almost occupied the entire coast from Zhejiang to Guangxi Province. We found the increased habitats caused by coupling effects mainly distributed in the landward side of Guangxi and Guangdong and by shared effects were very small. Moreover, the decreased habitats caused by pure effects of sea level rise and climate change, which were very small, mainly distributed in Zhejiang coast and Pearl River Estuary, respectively.

| D ISCUSS I ON
Earlier studies used the rules provided by MaxEnt (percent contribution, permutation importance, and jackknife test) to determine the variable importance (Fand et al., 2014;Liu et al., 2018;Saatchi et al., 2008;Smart et al., 2012;Yao et al., 2016). Given interactions among variables were unavoidable, variable importance should be interpreted with caution when using these traditional methods (Parisien & Moritz, 2009). However, the global sensitivity analysis can reveal the importance of the main and total effects of different variables with considering interactions among variables (Haaker & Verheijen, 2004;Liu et al.,2019). Our research demonstrated that Bio02 and Elevation were the most important variables in controlling the distribution of S. alterniflora on the northern coast, whether main or total effects. The finding was consistent with previous studies (Kirwan et al., 2010;Liu, 2018;Priest, 2011). However, Elevation is by far the most important variable on the southern coast due to the width of tidal flat was narrow and the altitude has a relatively high limitation for the expansion of S. alterniflora (Gao et al., 2014). Bio02 and Elevation showed strong interactions among variables. An explanation for this might be that altitude can directly affect climatic factors such as temperature and precipitation and in turn climatic factors affect species distribution at different altitudes (Crosby et al., 2017;Idaszkin & Bortolus, 2011;Marangoni & Costa, 2012;Zhao et al., 2015 Our studies also showed that habitats change was influenced not only by pure effects of climate change and sea level rise but also by shared and coupling effects of their interactions, which is similar with the previous studies that habitats change was influenced by their interactions in a complex manner (Hering et al., 2009;Milad et al., 2011;Reyer et al., 2013;Wu, 2017 Overall, our findings illustrated that the distribution of S. alterniflora was controlled not only by the pure effects of climate changes and sea level rise, but also by the shared and coupling effects caused by their interactions in different regions. Thus, climate changes, sea level rise, and their interactions should be taken into consideration for robust predictions of the spatial distribution patterns of S. alterniflora. It will provide more scientific and reasonable suggestions for preventing and controlling the invasion of S. alterniflora. Although ecological niche modeling (MaxEnt) is a superior technology for modeling the potential distribution of species, it has several limitations including its uncertainty and transferability (Phillips et al., 2006;Swanson et al., 2013). Given the model's uncertainty, our research was built on 10-fold cross-validation and multiple threshold rules, together with its high accuracy, and all F I G U R E 5 The spatial distribution of the changed habitats of spartina alterniflora caused by pure, shared, and coupling effects of sea level rise and climate change. (a) the northern Chinese coast (b) the southern Chinese coast of them supported the reliability of the results obtained (Elith & Yates, 2015;Radosavljevic et al., 2013). MaxEnt assumes that species will not exhibit phenotypic adaptation to new environmental conditions (Hernandez et al., 2006). Our model did not account for species dispersal, while the seeds of S. alterniflora can spread over long distances by wind and waves. Thus, further work is needed to combine the dispersal of S. alterniflora to better predict its actual distribution. Furthermore, we only assumed that the average sea level will rise by 1 meter without considering the spatial heterogeneity of sea level rise. Abiotic environmental factors such as interspecies competition and ecosystem dynamics could also influence S. alterniflora's survival and colonization success (Woolfolk, Wasson, 2013;Garner et al., 2015). Therefore, further studies require considering the effects of biological factors such as species dispersal, competition, the spatial patterns of sea level rise in different regions.

ACK N OWLED G M ENTS
This research has been supported by National Natural Science

AUTH O R CO NTR I B UTI O N S
Haibo Gong formed the original idea and wrote the original manuscript; Huiyu Liu offered valuable comments and was responsible for the manuscript revisions; FuSheng Jiao created figures and tables; Zhenshan Lin and Xiaojuan Xu analyzed the data.

DATA ACCE SS I B I LIT Y
The datasets are available at http://data.gbif.org/, www.worldclim.