Rhododendron diversity patterns and priority conservation areas in China

2.1 Species data

China forms the study area. Records on Rhododendron presence were collected from seven Chinese herbaria and botanical museums (for more details, see Yu et al., 2015). As high locational accuracy is required for studying plant species distribution, all records presenting only a general description of the location (e.g. mentioning only a county or a mountain) were excluded. Our resulting dataset, covering 406 species, comprises 13,126 geo-referenced records, with each record having a spatial uncertainty of less than 1 km. We chose a grid cell of 10 × 10 km to model Rhododendron diversity as we attempt to provide as much detail as possible about the spatial patterns of Rhododendron diversity for conservation planning in order to efficiently allocate scarce resources. A grid cell of 10 × 10 km matches the input data used in this study (see the following section for details). All points of the same species fall into a 10 × 10 km grid cell are considered as one occurrence, and those species with more than 10 occurrences were retained for further analysis. This approach narrowed the dataset down to 212 species with 9,360 occurrence records. The geographic coordinates of each record were projected onto an Albers equal-area conic conformal coordinate system to avoid the latitudinal biases of geographic coordinate systems (for the distribution map, see Figure 1). We are aware that the sample distribution may influence the output of the distribution modelling (Merow et al., 2014; Phillips et al., 2009). But after an initial test involving the manipulation of the background data, and recognizing the long-term and intensive data collection efforts by the various herbaria, we consider sampling bias to be unlikely.

2.2 Environmental variables

We selected a comprehensive set of environmental variables that were considered to be the main factors influencing the distribution of plant species (Dirnbock, Dullinger, & Grabherr, 2003; Körner, 1999). To deal with collinearity, we only considered variables with a variance inflation factor (VIF) less than 8. The climatic variables used in this study were as follows: isothermality (the ratio of the mean diurnal temperature range to the annual temperature range, indicating whether daily or seasonal temperature fluctuations are more dominant), temperature seasonality, annual precipitation and precipitation of driest month, all obtained from the WorldClim database (1950–2000) at a resolution of 5 arc minutes (Hijmans, Cameron, Parra, Jones, & Jarvis, 2005). As measures of habitat heterogeneity, we used elevation range and solar radiation, which were derived from a digital elevation model (USGS GTOPO 30 https://lta.cr.usgs.gov/GTOPO30) with a resolution of 30 arc seconds. In addition, normalized difference vegetation index (NDVI) and potential evapotranspiration (PET), which have both proved useful in predicting plant species distribution and richness (Cramer & Verboom, 2017; Williams et al., 2009), were also utilized. The NDVI data (8 km resolution) were derived from the third-generation Global Inventory Modeling and Mapping Studies (GIMMS 3g) product (http://glam1.gsfc.nasa.gov/), and averaged annually over the years 1982 to 2003. The PET data (at 30 arc seconds) were derived from the Consortium for Spatial Information (http://www.cgiar-csi.org/data/global-aridity-and-pet-database). The environmental layers were reprojected into the same Albers equal-area conic conformal coordinate system as the species data using R 3.1.3 (R Core Team, 2015) and aggregated to 10 km grid cells.

2.3 Spatially explicit species assemblage modelling (SESAM)

Following the method proposed by Guisan and Rahbek (2011), we used four steps and two evaluation indices in our implementation of SESAM.

2.4 Geographical pattern of species richness, β-diversity and weighted endemism

Rhododendron species richness was determined by the total number of species in each geographical unit. Many indices have been used to calculate β-diversity (Tuomisto, 2010). We used Simpson's beta (β_sim, Equation 1) index, which corrects for differences in species richness between sites (Joger et al., 2014; Zhang, Slik, & Ma, 2016). The pattern of β-diversity was estimated using a moving-window method (Lennon, Beale, Reid, Kent, & Pakeman, 2011). Based on previous β-diversity studies in China (Wang, Fang, Tang, & Shi, 2012), we set the window size at r = 50 km. To avoid a possible bias in β-diversity estimates near the border, grid cells covered by less than half of the moving window were excluded. To assess the β-diversity pattern at grid scale, we calculated β-diversity for each cell as the mean of the dissimilarity values between this focal cell and each of its adjacent cells.

(1)

where A is the number of species found in both cells i and j; B is the number of unique species in cell i; and C is the number of unique species in cell j. The range of β_sim runs from 0 to 1, with 1 denoting no species in common between two grid cells (complete dissimilarity) and 0 denoting that two grid cells contain identical species (no dissimilarity). Weighted endemism was calculated by first weighting the presence cells of each species by the inverse of its predicted range size (frequency of occurrence, i.e. number of 10 × 10 km grid cell where species occurred), and then calculating the sum of these scores cell by cell (Williams et al., 2009; Herkt et al., 2016, Equation 2).

(2)

where t is a species in {T}, {T} is a set of species found in the study area and R_t is the range size of species t. For consistency, the patterns of species richness and weighted endemism were also generated with a moving window (r = 50 km). All the spatial analyses were conducted using Biodiverse 1.1 (Laffan, Lubarsky, & Rosauer, 2010).

2.5 Model evaluation

Species richness, β-diversity and weighted endemism predicted from SESAM were compared with the observed values using the Spearman rank correlation. To evaluate the species composition output of SESAM, a confusion matrix was used in which all species were classified into either true positive (TP), true negative (TN), false positive (FP) or false negative (FN) for every grid cell, with the total number of species (SP) as defined in step 1. Next, the assemble prediction success (see Equation 3) and Sørensen index (see Equation 4) were calculated for each grid cell, based on this matrix.

(3)

(4)

2.6 Identification of Rhododendron diversity hotspots

Various criteria have been used to quantify biodiversity hotspots in previous studies, for example, top 5% of record-containing grid cells (Prendergast, Quinn, Lawton, Eversham, & Gibbons, 1993), 0.5% of all plant species worldwide (Myers et al., 2000), top 2.5% of grid cells regarding species richness (Orme et al., 2005) and 5% of total land area with highest biodiversity (Huang et al., 2016). In this study, given that we were evaluating a single genus with three diversity metrics and a grid cell of 10 × 10 km, the above criteria may not be suitable as the concept was developed for multiple genera and coarser grid cells (Pardo et al., 2017; Zhao, Li, Liu, & Qin, 2016). Therefore, we identified hotspots by considering the top 1%, 5%, 25% and 50% of record-containing 10 km cells regarding species richness, β-diversity and weighted endemism separately, as well as jointly (i.e. spatial congruence, quantified by counting the cells that belong to hotspots of species richness, β-diversity and weighted endemism at the same time).

2.7 Analysis of the conservation gaps

The conservation gap areas were identified by overlapping the three metrics with the distribution of Chinese nature reserves (Figure 2). To quantify Rhododendron's conservation status, we calculated the proportion of hotspots protected by nature reserves relative to the total area of hotspots, based on different levels (the top 1%, 5%, 25% and 50%) of diversity for defining the hotspots. Of the 212 Rhododendron species, 43 species are “threatened” species as classified under the IUCN Red List Categories of Critically Endangered, Endangered and Vulnerable (Gibbs et al., 2011; IUCN, 2001). Given the high conservation value of threatened species, we repeated the calculations for the threatened Rhododendron species. This procedure was implemented in ArcGIS 10.2 (ESRI Inc., Redwoods, USA) and r 3.1.3 (R Core Team, 2015).

3.1 Distribution patterns of Rhododendron diversity

The SESAM integrated from the individual SDMs has high AUC (0.98 ± 0.01), TSS (0.94 ± 0.03) and Kappa (0.76 ± 0.05) scores. The correlation between the three predicted diversity metrics (i.e. species richness, β-diversity and weighted endemism) and the observed ones is 0.96, 0.85 and 0.71, respectively (p < .05, Fig. S1). The two composite prediction indices vary from 0.6 to 0.9 for the prediction success, from 0.1 to 0.9 for the Sørensen index, while rising across all grid cells (Fig. S2).

Overall, Rhododendron species richness, β-diversity and weighted endemism (Figure 3a–c) display a similar trend across China, decreasing from the south-west (Mt Hengduan) to the north-east (Mt Daxinganling), mainly along the mountain ranges. All the area with the highest species richness (>80), β-diversity (>0.96) and weighted endemism (>2.5) are located in south-western China (i.e. north-western Yunnan), while the large area in north-eastern China exhibits the lowest diversity (species richness <7, β-diversity <0.68 and weighted endemism <0.03). In the north-west, especially in Xinjiang and Ningxia province, Rhododendron is not expected to occur at all.

3.2 Distribution patterns of Rhododendron hotspots

When defining the top 1% of grid cells as hotspots, the hotspots identified by the three diversity metrics, separately (circle 1 in Figure 3a–c, also see Fig. 4a1, b1 and c1) as well as combined (Figure 4d1), are all concentrated in Gongshan County (north-western Yunnan). When the hotspot defining level is expanded to the top 5% of grid cells, the hotspots of species richness include three more patches (i.e. Deqin, Lijiang and Dali) in north-western Yunnan (circles 2, 3 and 4 in Figure 3a, and see also 4a2), while another two more locations (i.e. Motuo and Luding, circles 5 and 6 in Figure 3b, and see also 4b2) are detected using the β-diversity metric. Hotspots of weighted endemism are distributed sporadically in north-western Yunnan, south-eastern Tibet, south-western Sichuan and also a small patch in western Hubei (Figure 3c, and see also 4c2). But the congruent hotspots are only found in north-western Yunnan (Figure 4d2). When defining the top 25% grid cells as hotspots, south-eastern Tibet and south-western Sichuan emerge as species rich (Figure 4a3), while more spots, including southern Chongqing, south-western Jiangxi and northern Guizhou, are detected by β-diversity (Figure 4b3). North-western Chongqing (Mt Daba) is a relatively large hotspot detected by weighted endemism (Figure 4c3), but the congruent hotspots are mainly at Mt Hengduan and in south-western Sichuan (Figure 4d3). When defining the top 50% grid cells as hotspots, more hotspots emerge in south-eastern China (Figure 4a4, b4 and c4). The congruent hotspots of the three metrics then cover about 133,200 km², which accounts for 1.4% of the total land area of China. Meanwhile, the congruent hotspots account for 56.4% of the area defined by the three metrics (Figure 5a), and the difference in overlap between these three metrics is largest at this level (Figure 5b), which indicates that the most unique hotspots based on either species richness, β-diversity or weighted endemism can be found at the 50% level. Therefore, we consider the spatial congruence (common area) of the richest 50% grid cells of three metrics as the ideal threshold for defining Rhododendrons hotspots. In total, 12 hotspots are identified (Figure 6), which are: (1) southern Chongqing (Mt Jinfo), (2) south-eastern Tibet, (3) Mt Hengduan (Yunnan) and north-eastern Yunnan, (4) south-western Sichuan (Mt Shaluli, Mt Daxue and Mt Qionglai), (5) northern Guangdong (Mt Danxia), (6) southern Zhejiang, (7) north-eastern Guizhou (Mt Fanjing), (8) western Hubei (Mt Daba), (9) south-eastern Jiangxi, (10) south-western Hubei and northern Hunan, (11) northern Guangxi (Mt Maoer and Mt Yuanbao) and (12) northern Fujian.

3.3 Diversity patterns and hotspots of threatened Rhododendrons

Patterns of species richness, β-diversity and weighted endemism for threatened Rhododendron species (Figure 7) are consistent with the diversity patterns of overall Rhododendron species. Following the hotspot definition of overall Rhododendron species, the spatial congruence of the top 50% of grid cells of the three diversity metrics, we found that the hotspots of threatened Rhododendron species are largely overlapping (86.3%) with hotspots of overall Rhododendron species. The five hotspots for threatened Rhododendron species are (Figure 8): southern Chongqing (Mt Jinfo), south-eastern Tibet, north-western Yunnan, south-western Sichuan and northern Guangdong.

3.4 Priority areas for Rhododendron conservation

The hotspots of Rhododendrons are only partly covered by the 2139 nature reserves in China (Figures 4 and 6). The hotspots and the nature reserves have 38,106 km² in common, which accounts for 2.3% of the total area of China's nature reserves. The proportion of protected hotspots by nature reserves (Figure 9) shows that about 96.7% of the congruent hotspot areas at the top 1% level are protected, although this ratio decreases to 52.9% when the congruent hotspot threshold is extended to the 5% level. In general, the protection ratio decreases with a decrease in the threshold defining the hotspots. For the top 25% of congruent hotspots, the average protection ratio is only approximately 28.3%. Given that these hotspots are mainly located in 12 provinces, we present the proportions of the hotspots that are protected for threatened and overall Rhododendron species in the 12 provinces (Table 1). It should be noted that southern Chongqing is identified as a hotspot for both threatened and overall Rhododendron species. However, only 23.1% of area where threatened Rhododendron species grow are protected. In Tibet, which has also been identified as hotspot area for both threatened and overall Rhododendron species, 27.4% of threatened and 36.3% of overall Rhododendrons are protected. In the largest hotspot (i.e. north-western Yunnan), which harbours more than half of the threatened Rhododendron species in China, the protection ratios for threatened and overall Rhododendron species are 39.6% and 31.8%, respectively. In the second largest hotspot (i.e. south-western Sichuan), 40.4% and 20.2% of, respectively, the threatened and overall Rhododendrons are protected. The two small hotspot areas with least numbers of Rhododendron species, namely southern Zhejiang (hotspot 6 in Figure 8, about 500 km²) and north-eastern Guizhou (hotspot 7 in Figure 8, about 849 km²), are barely protected at all.

4.1 Patterns, potential hotspots and the conservation status of Rhododendrons in China

Rhododendrons form a keystone element of montane ecosystems in the alpine and subalpine zones of south-eastern Asia. Rhododendrons growing at higher altitudes are sensitive to disturbance by natural factors, such as climate change, landslides and forest fires. Rhododendrons that grow at lower altitudes are exposed to threats from human activity, such as the rising demand for agricultural land, road construction and the growing tourism industry (Ma et al., 2014). To the best of our knowledge, this is the first comprehensive study of Rhododendron diversity patterns across all of China in terms of richness, composition and rarity, thereby providing a foundation for decision-making regarding the conservation of Rhododendrons in China. The high diversity of Rhododendron species in the mountains (vs. the low diversity in the plains, basins and deserts) is consistent with the Chinese woody plant diversity pattern (Wang, Fang, Tang, & Lin, 2011) and endemic seed plant species diversity patterns (Huang, Lu, Huang, & Ma, 2015). The species richness pattern of higher plants tends to follow the trend of larger species numbers in the south and lower numbers in the north of China (Zhao et al., 2016). The 12 Rhododendron hotspots identified in China in this study are mainly located in southern China. Seven of these are considered to be centres of endemism of Chinese plant taxa (Lopez-Pujol, Zhang, Sun, Ying, & Ge, 2011), and eight are considered to be hotspots of Chinese threatened plant species (Zhang & Ma, 2008). All 12 hotspots have been identified as hotspots of Chinese endemic seed plants (Huang et al., 2016). Thus, to some extent, the Rhododendron hotspots represent plant diversity centres in China. However, the 12 hotspots identified in our study are only partly situated in nature reserves. The proportion of the Rhododendron hotspots protected by nature reserves in our study is low for both the small hotspots (southern Chongqing and southern Zhejiang, middle to low elevation regions) and the large hotspots (north-western Yunnan and south-western Sichuan, both high-elevation regions). North-western Yunnan has been recognized as the origin and distribution centre of Rhododendron species (Li et al., 2013). Consequently, we consider conservation of this genus to be of a high priority. The hotspots of threatened Rhododendron species, especially southern Chongqing, should be considered very high-priority areas (priority level I in Table 1) for conservation, and the remaining hotspots, especially southern Zhejiang and north-eastern Guizhou, should also be considered for conservation reasonably soon (priority level II in Table 1). Attention should also be paid to the hotspots in south-eastern China, which are located in densely populated areas and are seriously affected by land-use change (Yang, Ma, & Kreft, 2014), as to the well-known hotspots at Mt Hengduan and southern Chongqing.

4.2 Influence of diversity metrics and spatial scales on hotspots identification

Hotspot locations may differ depending on how the hotspot is defined. The selected diversity metrics, the level of spatial congruence of diversity metrics and spatial scale (i.e. the grid cell used in the analysis) are all critical factors that may lead to diverging results (Ceballos & Ehrlich, 2006; Marsh et al., 2010; Zhao et al., 2016). Various biodiversity metrics, including species richness, weighted endemism, phylogenetic diversity and biogeographically weighted evolutionary distinctiveness, have been used to portray species diversity, define hotspots and select priority areas for conservation (Huang et al., 2016; Pardo et al., 2017). Species richness is the most widely used metric, while weighted endemism has also gained much attention recently (Herkt et al., 2016; Huang et al., 2016). However, β-diversity (the dissimilarity of species composition between two sites) has not been commonly used to identify biodiversity hotspots and has attracted less attention than species richness (also called α-diversity on a local scale or γ-diversity on a larger scale). McKnight et al. (2007) demonstrated that β-diversity provides critical information for conservation planning when striving to represent biodiversity within practical constraints such as an area or coast. Marsh et al. (2010) used both species richness and β-diversity to model bird and butterfly diversity on the Comoro Islands, and concluded that species richness and β-diversity together provide comprehensive understanding when assessing optimal reserve locations. Socolar, Gilroy, Kunin, and Edwards (2016) concluded that understanding β-diversity is essential for protecting regional diversity and directly assists conservation planning. Our results confirm that species richness, β-diversity and weighted endemism describe different attributes of species diversity. Weighted endemism can be of value identifying “rarity hotspots” of species. Conservationists are interested in protecting areas with a high number of endemic species, which have a narrow distribution to a particular area (Crisp, Laffan, Linder, & Monro, 2001). And β-diversity facilitates the optimal spatial arrangement of conservation areas, capturing variation in species assemblages, as well as underlying environmental heterogeneity necessary for long-term persistence (McKnight et al., 2007). Therefore, we advocate that besides species richness, β-diversity and weighted endemism are employed for identifying hotspots, to provide a more comprehensive assessment of species diversity, consequently leading to more informed conservation prioritization. Hotspots of species richness, threatened species and endemism do, however, not always totally correspond (Orme et al., 2005), although our study did not encounter this problem. We did notice that there are few quantitative studies on the spatial distribution of threatened species and thus advocate emphasizing the quantitative assessment of threatened species in a diversity and hotspots study.

The importance of the chosen level of congruence in detecting hotspots has been clearly demonstrated in Figure 4. Based on the top 1–5% of grid cells of species richness, β-diversity and weighted endemism, only one hotspot emerged, namely Mt Hengduan, which is considered a global biodiversity hotspot (Myers et al., 2000). At this level, the remaining 11 potential hotspots would have gone unnoticed in the extremely diverse Himalayan region. Most hotspots are only unveiled when the “hotspots defining level” is expanded. The common area of the top 50% grid cells of species richness, β-diversity and weighted endemism was used as threshold to identify the Rhododendron hotspots, because, based on these three metrics, this level identified most hotspots. In addition, our hotspots are defined based on “record-containing grid cells.” This leads to the issue of grid cell size, which is important for hotspots identification in ecological studies (Pardo et al., 2017; Zhao et al., 2016). In continental-wide and global biodiversity studies, usually a grid cell of 100 × 100 km or 50 × 50 km is used for data analysis (Orme et al., 2005; Wang et al., 2011; Huang et al., 2015), but in contrast to use a 100 × 100 km grid, our 10 × 10 km grid leads to much more detailed patterns. Therefore, we suggest that the hotspots and priority area selection should take diversity metrics, the level of congruence and the basic grid cell size for data analysis into account, as well as comprehensively consider the specific circumstances of the region and the taxon involved.