Distribution data of P. wightiana were sourced from field surveys and the Chinese Virtual Herbarium (CVH, http://www.cvh.ac.cn). Based on the National Natural Science Foundation of China (No.4156010383), 45 natural populations of P. wightiana were obtained by extensive field investigation. Additionally, complete distribution information from CVH were first confirmed and then recorded. If there was a detailed collection site, the longitude and latitude were located using Google Earth with reference to the recorded habitat, altitude, along with other information. Voucher photos were checked for records sourced from CVH and NSII (National Specimen Information Infrastructure) to confirm that the species were correctly identified. In total, 48 collection entries from herbarium research were obtained after removing repetitive entries and identification errors.

To prevent model over-fitting caused by data repetition and spatial autocorrelation, as well as to associate model error to the results, the buffer method was used to screen the obtained data. The spatial resolution of environmental variables was 2.5 arc-minutes and spatially coincident data points within 3 km of each other were also eliminated (Wang et al. 2017, Guo et al. 2019). Finally, a total of 91 filtered occurrence records were obtained from CHV and NSII, along with our own collections, which well covered the distribution range of P. wightiana (Fig. 1, Suppl. material 1)

Figure 1.  

Geographic distribution sample points of Parnassia wightiana.

The bioclimatic variables (Hijmans et al. 2005), downloaded from the WorldClim database (http://www.worldclim.org), included nineteen climatic factors representing historical periods, i.e. LIG, LGM, MH, along with Current and Future periods. Future climate selection RCP (Representative Concentration Pathways) 6.0 represents the greenhouse gas emission peak scenario for projections over the distribution area (Wang et al. 2017). The common climate model CCSM4 (Community Climate System Model) was used to source the environmental data for each period and the 2.5 arc-minutes spatial resolution was adopted. The geographical coordinates were unified as GCS_WGS_1984. The ArcGIS 10.4 extract by mask tool was used to cut and extract the environmental variable data, retaining the required range of data and, finally, to convert to ASCII format. A 1:4 million scale provincial administration map was downloaded from the China's National Basic Geographic Information System (http://nfgis.nsdi.gov.cn) as the analysis base map.

High correlation and collinearity between bioclimatic variables can easily lead to the over-fitting of the model, thus affecting the accuracy of the resulting predictions, so SPSS 26 was utilised to perform principal component analysis on all bioclimatic variables (Yang et al. 2013). Bioclimatic variables corresponding to the species distribution points were extracted and then imported into MaxEnt v.3.4.1 together with the species data for suitability prediction, where the climate variables greater than 1% in the model prediction results were retained (Shi et al. 2021). To eliminate multicollinearity effects in the parameter estimates of the SDM, variables with Pearson’s |R| ≥ 0.85 were excluded (Yusup et al. 2018) and only six variables were selected as climatic predictors to model the past, current and future climatically-suitable areas of P. wightiana. These six bioclimatic variables, namely Bio2, Bio4, Bio7, Bio11, Bio12 and Bio14, were retained for the construction of the distribution prediction model (Table 1).

The sorted distribution point data and the screened bioclimatic variable data were imported into MaxEnt 3.4.1 and the bioclimatic variables were evaluated by the Jackknife test. The models obtained were calibrated using 75% of the available records for each species as training (calibration) data and the remaining 25% were used for model validation as test data. The Bootstrap method was implemented with 10 repeats, a maximum of 5000 iterations and default selected parameters (Yan et al. 2020).

Model performance was evaluated by calculating the Area Under the Receiver Operator Curve (AUC), where models with AUC values larger than 0.7 were considered satisfactory for our study (Wiley et al. 2003, Phillips and Dudik 2008). MaxEnt is a multivariate approach that estimates the distribution of a species by finding the probability distribution of maximum entropy, subject to constraints representing our incomplete information about the distribution. It is generally understood that AUC < 0.7 indicates low accuracy of the model, prediction results can be adopted when AUC is 0.7–0.9 and AUC > 0.9 indicates that the prediction results are very accurate, which can be used for subsequent analysis.

A MaxEnt model was used to predict the distribution of P. wightiana in different periods, which were then visualised using ArcGIS. The prediction of the MaxEnt model, based on the current climate data, is very consistent with the actual distribution area and the AUC values of each period are greater than 0.9, indicating that the model has a good predicting ability (Fig. 2).

Figure 2.  

The results of the AUC in developing Parnassia wightiana habitat suitability model.

Note:*a: LIG, b: LGM, c: MH, d: Current, e: Future.

To distinguish unsuitable habitat from suitable habitat, a reclassification of the probability maps was performed using a threshold, which establishes the minimum level below which a given distribution should be excluded. The threshold of automatic generation, based on the model, is 0.277, that is, the range of 0-0.277 represents the non-suitable area of P. wightiana. Suitable areas are divided into the following grades: low suitable areas (0.277-0.518), medium suitable areas (0.518-0.759) and high suitable areas (0.759-1). Total counts of the grid numbers of each grade were used to calculate the area of each suitable region (Table 2).

The most influential environmental variables (Table 1) used by the Maxent model for the best model performance were Bio2, Bio4, Bio7, Bio11, Bio12 and Bio14. The contribution rates of annual precipitation (Bio12), annual average temperature range (Bio7) and mean temperature of the warmest quarter (Bio10) in each period were ranked in the top three and constituted more than 17% of the total contribution. The six bioclimatic variables, selected here, are known to influence the distribution and physiological performance of plant species (Table 3). Parnassia wightiana is sensitive to habitat and climatic changes and requires a specific forest habitat that is rich in water and adequate temperature. Results from the Jackknife test indicate that precipitation was the most influential factor.

In the present study, we found that the MaxEnt model was able to provide robust results with small, sparse and irregularly sampled data and our results are visualised using ArcGIS. The operation was simple and intuitive and the sample requirement was small and other previous studies have also highlighted the great performance of similar methods. For example, De Siqueira et al. (2009) identified a rare plant species that had already been declared extinct through ecological niche simulation. For many species, in fact, the sample field data are insufficient to characterise the geographic distribution of species in the study area and, often, must be supplemented by digital specimen information. However, the distribution records of some specimens must be confirmed by Google Earth due to the lack of detailed latitude and longitude information. Some distribution sites have been collected over long periods and their environment has changed over time. Using these distribution data for simulation often weakens the model's niche definition, resulting in a decline in the simulation ability and accuracy of the distribution model.

In this study, the environmental variables used in the MaxEnt model are all climate factors. Nineteen climatic variables, based on temperature and rainfall, were chosen according to the different needs of computing performance. However, there are many factors in addition to climate that affect the distribution of species, such as interspecific interactions, microtopography and local microclimates. For example, the uplift of the Qinghai-Tibet Plateau, along with the environmental destruction caused by human activities may together lead to species extinction (Montemayor et al. 2015). Therefore, the MaxEnt model represents the theoretical maximum possible species distribution and the suitable areas are often shown to be much wider than the actual areas occupied. The application of models, based on the principle of maximum entropy for predicting species distributions according to the extent of suitable area, represents a useful tool, both to define the most important environmental variables responsible for current distributions and to provide useful information for their conservation.

The predicted distribution of P. wightiana in each period showed that the suitable areas were concentrated in the mountainous areas of south-western China, mainly in the Himalayan, Hengduan and Qinling Mountains. According to the existing collection along with specimen records, Daming Mountain in Wuming District in Guangxi represents the southern distributional limit, the Weiyuan Weihe River in Gansu represents the northern, Huanggang Mountain in Jiangxi the eastward and Jipu Village in Jilong Town in Tibet constitutes the westward limit. Additionally, Nepal, Bhutan, northern Myanmar, Thailand, Sikkim and northeast India are also distribution areas for P. wightiana (Shu 2017). Overall, the distribution range (23°-35°N, 85°-118°E) spans over three endemic centres, one in the Hengduan Mountains, another in Central China and finally in the Lingnan Region of China, showing an obvious discontinuous distribution.

Based on the results of the MaxEnt model, the key factors affecting species geographical distribution are not completely consistent. For example, a previous study on Stipa purpurea in the Tibetan Plateau found that precipitation was a key factor affecting its distribution (Hu et al. 2015). For Rosa roxburghii, temperature displayed the greatest influence on the future distribution in a simulation study (Fan et al. 2021). The results of our analysis indicated that annual precipitation was the most important climatic factor affecting the distribution of P. wightiana. The species is mainly distributed along the edge of the Sichuan Basin and is confined to mountaintops. Therefore, average annual temperature and the average temperature in the warmest season are the key factors affecting the distribution of P. wightiana.

Understanding species distributional patterns is a fundamental question in biogeography and conservation biology. The middle and high suitable areas of P. wightiana were mainly located in the Hengduan Mountains, Yunnan-Guizhou Plateau, Qinling Mountains and Daba Mountain at high altitudes. During the Last Glacial Maximum, as temperatures gradually began to decline, some inland areas of Sichuan Basin began to transform into suitable areas. The Weihe River source in Gansu Province was a residual area of the Qinling Mountains and also began to transform into a medium suitable area. For the cold-adapted P. wightiana, the decline in temperature makes the environment more suitable for survival, most likely during the glacial expansion and postglacial contraction.

The study of potential plant refuges during the Quaternary glaciation is of great significance for understanding current plant distribution patterns along with future evolution. A MaxEnt model was employed to simulate the distribution of highly suitable areas over different time periods, allowing for the inference of potential refugia (Chan et al. 2011, Bai and Zhang 2014). The results of our analysis suggest that the Hengduan Mountains, Yunnan-Guizhou Plateau and Qinling Mountains represent the core areas of the potential distribution of P. wightiana, which provide more suitable habitats for survival than other areas. These findings may, therefore, provide useful information for increased understanding of the origin and differentiation of northern temperate species.

In this study, the distribution range of P. wightiana in the future was predicted. The results showed that the suitable area of P. wightiana would continue to shrink in the next few decades. Although the highly suitable area remained stable, the prospect of population development was not optimistic. In view of the small size of wild populations and serious human disturbance, we put forward some suggestions: (1) in situ conservation of the known P. wightiana community, minimise the adverse effects of human activities on species survival; (2) at present, the research on biological characteristics, artificial cultivation methods and genetic shape improvement of P. wightiana is relatively weak and systematic research in this area should be strengthened; (3) artificial harvesting and planting can be carried out around some populations with less human activity to maximise the size of their wild populations.