Climate change and landscape-use patterns influence recent past distribution of giant pandas

Abstract

Climate change is one of the most pervasive threats to biodiversity globally, yet the influence of climate relative to other drivers of species depletion and range contraction remain difficult to disentangle. Here, we examine climatic and non-climatic correlates of giant panda (Ailuropoda melanoleuca) distribution using a large-scale 30 year dataset to evaluate whether a changing climate has already influenced panda distribution. We document several climatic patterns, including increasing temperatures, and alterations to seasonal temperature and precipitation. We found that while climatic factors were the most influential predictors of panda distribution, their importance diminished over time, while landscape variables have become relatively more influential. We conclude that the panda's distribution has been influenced by changing climate, but conservation intervention to manage habitat is working to increasingly offset these negative consequences.

1. Introduction

The climate crisis challenges species persistence, alters species' geographical ranges, threatens global diversity and has caused policy-makers to rewrite conservation management strategy. The pace of these climate-driven ecological changes is accelerating [1–3]. While efforts to curb global greenhouse gas emissions address the ultimate driver of these effects, conservation management requires tools that can be implemented on regional scales [4] and in the near-term.

To optimize climate model use for conservation policy and management, models need to be biologically realistic. However, climate-only models can yield misleading results [5]. Further, Anthropocene era influences extend beyond climatic change and may include large-scale changes in human land-use patterns. These changes may be negative, leading to a loss of key natural resources such as forest cover, or may be positive, where active management to protect and restore natural systems may improve habitat suitability for target species. These processes can at times conflict and be difficult to disentangle. The ecological scale over which data are collected and analysed is yet another factor with critical implications for resulting conclusions [5], requiring efforts to evaluate how the selected scale improves or impairs model performance.

As an iconic species and global symbol of conservation, giant panda populations have undergone pronounced human-driven range contractions over the past 3000 years and the remaining distribution is highly fragmented [6,7]. However, the past two decades have witnessed modest gains in panda numbers and range expansions that are largely attributable to the establishment of protected areas and the implementation of species protections [8,9]. These improvements in the availability of suitable habitat contributed to the recent downlisting on the International Union for Conservation of Nature Red List from Endangered to Vulnerable [9], yet the prospect of climate change is considered a major factor increasing the panda's vulnerability to extinction, primarily through its effects on the panda's primary food source—bamboo [7,8,10,11]. Thus, conservation policy and planning needs to be informed by these conflicting anthropogenic processes, with climate and habitat degradation inflicting harm on panda populations and habitat protection and restoration having positive influences. Climate change forecasts for panda distributions have yielded widely varying results with projected habitat loss ranging from 12 to 71% [10,12–16], making recommendations for managers and policy-makers difficult to implement. While these forecast models have provided useful dialogue informing conservation strategies for pandas, no one, to our knowledge, has yet taken a comprehensive look at what has happened to panda distributions under recent past changes in climate. Therefore, a retrospective analysis determining how changing climate and ecological landscapes already has influenced panda distribution will inform the debate on climate change impacts for the species and place climate impacts in the broader context of other threats and conservation measures.

Here, we use an unprecedented dataset of spatially explicit species presence records of the giant panda (Ailuropoda melanoleuca) and environmental data over the past 30 years from Sichuan province, China, to evaluate the role of climatic, ecological and anthropogenic factors in determining changing distributions of the species and investigate whether the importance of these factors has changed over time. To do so, we compared the predictive ability and temporal transferability of models based on only climate factors to those models which also include land-use and/or topographic factors, and explore how model differences manifest across different time periods and grid resolution. We also explore how the importance of various predictive variables has changed over time and across different grid sizes. According to our knowledge, our study is one of the first studies to assess the relative importance of climate change and landscape-use patterns in panda distribution and addressed whether climate change has influenced panda distribution.

2. Material and methods

(a). Study area and panda presence data

To date, four National Giant Panda Surveys have been completed to monitor panda status in China. For each survey, panda presence was determined via signs (e.g. faeces, fur and signs of foraging). Additional details on survey methods can be found in China's State Forestry Administration report [17]. This study was conducted in Sichuan province, China—home to about 75% of giant pandas (figure 1). The panda presence data during three time periods (TP1: 1985–1988, TP2: 1998–2002 and TP3: 2011–2014) was obtained from China's State Forestry Administration reports [17,18]. In total, we obtained 860, 3286 and 3428 presence records during TP1, TP2 and TP3, respectively. To investigate the effect of spatial scale on model performance, we created 5 km × 5 km, 10 km × 10 km and 15 km × 15 km grids and excluded duplicate records for each grid cell. During the procedure, we obtained 433 panda occurrences during TP1, 519 occurrences during TP2 and 602 occurrences during TP3 at 5 km × 5 km; 226 occurrences during TP1, 236 occurrences during TP2 and 258 occurrences during TP3 at 10 km × 10 km; and 147 occurrences during TP1, 142 occurrences during TP2 and 160 occurrences during TP3 at 15 km × 15 km.

Figure 1.

Distributional range of giant pandas in Sichuan province, China, during the past three decades. (a) Panda presence in the 2nd national survey (1985–1988; TP1), (b) panda presence in the 3rd national survey (1999–2001; TP2) and (c) panda presence in the 4th national survey (2011–2014; TP3). (Online version in colour.)

(b). Environmental data and variable calculations

For each time period, monthly temperature and precipitation measures for 1979–2018 were obtained from the China Meteorological Forcing Dataset [19], which provides monthly gridded data with a spatial resolution of 0.1°. Based on these data, we created 19 bioclimatic variables, following the procedure described by Hijmans et al. [20] using the biovars function in the R package dismo (https://CRAN.R-project.org/package=dismo). Each of the 19 bioclimatic variables was aggregated to each of the three scales (grid cells) and masked to include only the study area to match to the species distribution maps. All spatial analyses were conducted using ArcGIS 10.2 (ESRI, 2016).

The land-use data were obtained from multi-period remote sensing monitoring datasets of land use and land cover in China (CNLUCC; http://www.resdc.cn), which provides gridded data with a spatial resolution of 30 m. For analysis, we included the proportion of area covered by six different land-use types (farmland, forestland, grassland, water, residential land and other land). We selected land-use data collected in 1990, 2000 and 2010 for TP1, TP2 and TP3, respectively, because these time periods were nearly coincident with the timing of panda presence surveys, reflecting land-use patterns at the time of the survey. For each time period and grid resolution, we masked the land-use layers to the study area and calculated the proportion of each land-use type.

The digital elevation model (hereafter DEM) data were extracted from Earth Env-DEM90 dataset (http://www.earthenv.org/DEM), with 90 m spatial resolution. From this dataset, we extracted the elevation and calculated the slope and aspect for each 90 m × 90 m grid cell, then we calculated the mean elevation, mean aspect, mean slope and the range of elevation for each of the three grid cell resolutions (5 km × 5 km, 10 km × 10 km and 15 km × 15 km).

(c). Species distribution models and temporal transferability

A maximum entropy model (MaxEnt; [21]) was used to derive the relative importance of environmental factors correlated with panda presence during the 2nd (TP1), 3rd (TP2) and 4th (TP3) surveys. We chose MaxEnt because of its capacity to model species distribution using presence-only data. For each time period and grid resolution, we ran MaxEnt with seven models based on all possible combinations of the predictor variables. Model combinations included: (i) a full model that included all the climatic, land use and topographic predictor variables (CLT), (ii) a model that included the climatic and land-use predictor variables (CL), (iii) a model that included the climatic and topographic predictor variables (CT), and (iv) a model that included the land-use and topographic predictor variables (LT). Each of the remaining models included single predictor variables: (v) a climatic model (C), (vi) a land-use model (L), and (vii) a topographic model (T). To reduce the multicollinearity among environmental variables and ensure that selected variable sets were suitable for all nine datasets (three grid resolutions and three time periods), Pearson's correlation coefficients were used for variable selection. When two variables were highly correlated (≥0.70; [22]), we selected the variable with greater ecological relevance into subsequent modelling. Finally, we selected four climate variables, four land-use variables and two topographic variables for modelling species distribution. Climate variables included mean annual temperature (BIO1), temperature seasonality (BIO4), precipitation of driest month (BIO14) and precipitation seasonality (BIO15). Land-use variables included farmland (FAL), forestland (FOL), residential land (RL) and other land (OL). Topographic variables included mean aspect (MA) and the mean range of elevation (RE).

We then performed cross-validation on each model using a bootstrap approach, where random subsets of 80% of each dataset were for training data and the remaining 20% for testing model performance. At each cross-validation step, the predictive performance of each model was assessed by using the following two metrics: (i) the area under the receiver operating characteristics curve (AUC), which is a commonly used metric of model predictive performance and can range from 0.5 (predictive performance no better than random) to 1.0 (perfect predictive performance); and (ii) the point-biserial correlation (COR), which was used to quantify the degree of correlation between model predicted probabilities. In order to assess the relative importance of climate and non-climatic factors in determining panda distribution, for each covariate included in the full models, we calculated variable contribution as one minus the correlation between the predicted values and predictions where the variable under investigation is randomly permutated [23,24]. This procedure was repeated 10 times to produce predictions that were independent of the training data [25] for each of the above seven models across grid cell resolutions and time periods, generating a total of 630 models for comparisons.

For each spatial scale and time period, the models trained were projected to the datasets of the other two periods (hindcasting or forecasting), respectively. To evaluate temporal transferability of each model (i.e. the predictive ability of models beyond the temporal scope of data used for the training and testing of species distribution models (SDMs)), projections at each time period were tested using AUC as an accuracy metric, and then, a transferability index (TI) adapted from [26] for each model was calculated as follows:

$$\mathrm{TI}=\frac{1}{3}[f\hspace{0.17em}(T_A,\hspace{0.17em}{T}_B)+f\hspace{0.17em}(T_B,\hspace{0.17em}{T}_C)+f\hspace{0.17em}(T_A,\hspace{0.17em}{T}_C)],$$

where

$$\text{f}(T_A,\hspace{0.17em}{T}_B)=\frac{{\displaystyle \frac{1}{2}[(1-|({\mathrm{AUC}}_{T_A\to T_A}-{\mathrm{AUC}}_{T_A\to T_B}/0.5)|)+(1-|({\mathrm{AUC}}_{T_B\to T_B}-{\mathrm{AUC}}_{T_B\to T_A}/0.5)|)]}}{1+||({\mathrm{AUC}}_{T_A\to T_A}-{\mathrm{AUC}}_{T_A\to T_B}/0.5)|-|({\mathrm{AUC}}_{T_B\to T_B}-{\mathrm{AUC}}_{T_B\to T_A}/0.5)||},$$

and where ${\mathrm{AUC}}_{T_A,T_A}$ is the evaluation of the model fitted in T_A and evaluated on the same time period using data partitioning (internal evaluation measures) and ${\mathrm{AUC}}_{T_A,T_B}$ fits the model with data in T_A and evaluates it on data in T_B (external evaluation measures). The TI provides a measure of any reduction in model performance when going from internal to external evaluation, with TI values ranging between 0 and 1. A TI of 1 indicates no loss of predictive performance. Temporal transferability is especially germane for projecting changes in distribution owing to forecast climate change [27].

(d). Statistical analyses

Using Student's t-test, we evaluated whether significant changes in climatic and land-use variables at each grid cell resolution took place between time periods (TP1-TP2, TP2-TP3 and TP1-TP3). To assess the relative importance of different predictor sets in determining the distribution of the giant pandas, Wilcoxon sign rank tests were used to compare the performance of the various models based on the AUC values and COR values of the various models for each grid size and time period. Two–tailed tests were used for model comparisons between models based on non-overlapping sets of predictors, while one-tailed tests were used for comparisons of nested models.

Based on importance values obtained for each predictor variable (figure 2), three-way analyses of variance (hereafter ANOVA) were used to test whether the importance of predictor variables differed among time periods, grid sizes and type of predictor variables (EV1: climatic and non-climatic variable sets; EV2: temperature, precipitation, land-use and topographic variable sets). The average contribution values of variable sets were used as response variables, and environmental variable sets, grid sizes and time periods (TP1, TP2 and TP3), as well as their two- and three-way interactions as explanatory variables. We further tested the pairwise difference in variable importance between time periods, grid sizes and type of variable sets by performing post hoc pairwise comparison tests (Tukey's HSD). All analyses were performed using the importance values of both AUC and COR tests derived through internal evaluation. Given that the importance values of AUC and COR tests are closed, we only present the COR test results and the AUC test results are not shown. The multiple comparisons were run using the multcomp package [28] with the glht function in the R platform (v. 3.5.1; http://cran.r-project.org).

Figure 2.

Change in the prevalence of climatic and non-climatic factors in panda habitat as a function of time period and spatial scale. Average changes in each variable were calculated as the later period values minus the earlier period ones and time periods were compared with Student's t-tests. (a) Bio1 = mean annual temperature; (b) Bio4 = temperature seasonality; (c) Bio14 = precipitation of the driest month; (d) Bio15 = precipitation seasonality; (e) FAL = proportion of farmland area; (f) FOL = proportion of forest land area; (g) RL = proportion of residential land area; (h) OL = proportion of other land area. More statistical details are found in the electronic supplementary material, table S2. ***p < 0.001; **p < 0.01; *p < 0.05; n.s., not significant.

3. Results

(a). Changes in climatic and land-use variables

Our analysis of climatic and land-use variables during this 30 year time period documented significant increases in mean annual temperature (+0.71°C), reductions in the seasonality of temperature, reduced precipitation in the driest months and increased seasonality of precipitation patterns (figure 2; electronic supplementary material, table S3). Changes were not unidirectional across time for all measures: precipitation seasonality first decreased in TP2 before increasing in TP3 and precipitation in the driest month first increased in TP2 before decreasing in TP3, indicating unstable climatic patterns. We found little evidence for land-use changes, although residential land-use has increased (significant only at the 15 km² grid scale). Although farm area decreased and forest area increased, these land-use patterns did not change significantly. The temporal stability of land-use and topographic factors provide a relatively controlled backdrop from which to examine the changing degree to which climate is influencing panda distribution over time.

(b). Evaluating climate model performance as a function of variable inclusion and scale

Our findings indicate that the full model (including climate, landscape and topography; CLT) performed significantly better than all other six models, and the models including climatic covariates performed significantly better than those not including climate (table 1 and figure 3). These results indicate that all variable types we evaluated contributed to model performance, helping to explain panda distribution patterns, and that climate contributed more to model performance than other variables. Our internal evaluation analyses confirmed that spatial scale influenced model performance: on average the 5 km × 5 km scale performed best, followed by 10 km × 10 km, and 15 km × 15 km (table 1 and figure 3). All but the full model predictive performance of data aggregated into 15 km × 15 km grid cells was relatively low (AUC < 0.8). These findings indicate that aggregation of data for climate, landscape use and topography at smaller spatial scales increases the ability of models to predict panda distribution.

Table 1.

Model validation procedures using area under the receiver operating characteristic curve (AUC). (Mean AUC values are displayed as function of model sets based on climatic (C), land-use (L) and topographic (T) variables, time period (TP1: 1985–1988, TP2: 1998–2002, TP3: 2012–2014), and internal and external evaluation procedures.)

					external evaluation, EE
		internal evaluation, IE			forecast			hindcast
scale	models	TP1	TP2	TP3	TP1 → TP2	TP2 → TP3	TP1 → TP3	TP2 → TP1	TP3 → TP2	TP3 → TP1	TI
5 km × 5 km	CLT	0.884	0.885	0.868	0.752	0.748	0.789	0.804	0.762	0.752	0.752
	CL	0.869	0.869	0.85	0.740	0.732	0.781	0.802	0.752	0.738	0.724
	CT	0.86	0.85	0.835	0.719	0.710	0.746	0.771	0.720	0.712	0.711
	C	0.844	0.838	0.815	0.697	0.713	0.711	0.758	0.689	0.654	0.677
	LT	0.761	0.786	0.769	0.775	0.777	0.767	0.772	0.781	0.770	0.964
	L	0.727	0.757	0.734	0.743	0.740	0.736	0.730	0.745	0.726	0.973
	T	0.714	0.725	0.72	0.724	0.718	0.715	0.723	0.727	0.722	0.947
10 km × 10 km	CLT	0.868	0.859	0.824	0.722	0.727	0.756	0.775	0.773	0.746	0.795
	CL	0.835	0.823	0.788	0.707	0.691	0.741	0.754	0.750	0.734	0.757
	CT	0.827	0.816	0.776	0.704	0.701	0.73	0.751	0.744	0.712	0.778
	C	0.811	0.796	0.753	0.671	0.65	0.684	0.722	0.700	0.650	0.72
	LT	0.751	0.76	0.727	0.762	0.748	0.738	0.759	0.767	0.759	0.923
	L	0.721	0.729	0.69	0.720	0.699	0.697	0.707	0.724	0.711	0.934
	T	0.712	0.714	0.665	0.715	0.700	0.695	0.723	0.712	0.715	0.934
15 km × 15 km	CLT	0.834	0.836	0.794	0.748	0.708	0.735	0.756	0.759	0.750	0.882
	CL	0.78	0.78	0.733	0.724	0.685	0.711	0.751	0.744	0.731	0.824
	CT	0.785	0.774	0.743	0.744	0.694	0.725	0.755	0.719	0.730	0.857
	C	0.778	0.757	0.706	0.692	0.654	0.663	0.721	0.687	0.666	0.778
	LT	0.707	0.717	0.68	0.749	0.733	0.721	0.741	0.763	0.749	0.832
	L	0.675	0.683	0.648	0.706	0.668	0.664	0.681	0.704	0.683	0.861
	T	0.692	0.693	0.689	0.720	0.708	0.709	0.719	0.726	0.724	0.849

Figure 3.

Model performance among SDMs generated with different variable sets based on AUC (a–c) and COR (d–f). Comparisons between model performance were made using a Wilcoxon sign rank test. For each measurement, different superscript letters (a–g) indicate significant difference between models. For each time period and grid size, the full model performed significantly better than all other models. CLT: models constructed with climatic, land-use and topographic variables; CL: models constructed with climatic and land-use variables; CT: models constructed with climatic and topographic variables; C: models constructed with only climatic variables; LT: models constructed with land-use and topographic variables; L: models constructed with only land-use variables; T: models constructed with only topographic variables.

(c). Role of climatic and non-climatic factors, time period and spatial scale in determining panda distribution

Our analyses showed significant differences in the relative importance of predictor sets among different time periods, spatial scales and types of predictor sets (figure 4; electronic supplementary material, tables S4–S10). In pairwise comparisons between predictor sets, we found that in general terms the most important predictors of panda presence can be ranked from most to least important as follows: temperature-related (all spatial scales and time periods), precipitation-related (most consistently at smaller spatial scales), and topographic and land-use patterns (electronic supplementary material, table S10). There were a few departures from this general pattern, most notably topography was more important than precipitation at the two larger spatial scales and land-use was more important than topography at the smallest spatial scale.

Figure 4.

Relative importance of variable sets (indicated in top left corner of each graph) in determining panda distribution across three time periods. Significance determined with three-way ANOVA. More statistical details are found in the electronic supplementary material, Appendix S1 and tables S4 and S7. ***p < 0.001; ns, not significant.

However, temporal patterns for these variables alter conclusions, with some of the most influential predictor variables showing decreased importance through time. Although mean annual temperature is clearly on the rise in panda habitat, across the three time periods the relative importance of climatic factors in predicting panda presence has been decreasing across all scales, whereas non-climatic factors have increased (figure 4). Both the relative importance of climatic factors' predictive ability generally, and temperature-related factors more specifically, have decreased over time (electronic supplementary material, tables S5 and S8). From TP1 to TP3, the mean importance of climatic factors decreased by an average of 1.87% to 4.24% depending on spatial scale (electronic supplementary material, table S5) and the mean importance of temperature-related factors decreased by an average of 4.39% to 6.67% across spatial scales (electronic supplementary material, table S8). By contrast, the importance of land-use and topography have been increasing but not consistently at all spatial scales analysed. When comparing between the first and third time periods, however, land-use patterns have become significantly more influential at all spatial scales examined. These findings highlight the value of having a robust dataset for an individual species across relatively long temporal scales.

The sensitivity of model outcomes to spatial scales was apparent in that all results were dependent on the grid cell resolution (electronic supplementary material, tables S4, S6 and S9). In each time period, the average relative importance of all climate variables was positively correlated with grid cell resolution, while that of land-use variables and topographic variables were negatively correlated with grid cell resolution (figure 4). However, our external evaluation of model performance and subsequent calculation of a TI (table 1) demonstrated higher temporal transferability for analyses using 15 km × 15 km grid cells than those using 10 km × 10 km or 5 km × 5 km for all models. Indeed, at the smaller scales, models that included climate covariates had substantially lower temporal transferability. This indicates that modelling panda distribution at the spatial scale of 15 km × 15 km has low validity.

4. Discussion

A better understanding of how climatic and land-use changes have interacted to influence species distributions through time provides a platform for more informed conservation management strategies. Climatic variables are less under the control of managers and decision-makers at regional and local scales, but non-climatic factors such as land-use patterns and forest cover are subject to more local control, especially for protected areas. By mitigating against human land-use in panda habitat and encouraging forest recovery, managers can offset some of the negative effects of a changing climate, and our analyses indicate that these management actions should have larger positive impacts on panda distribution than they would have three decades ago. Our findings also inform methodological considerations for climate change analysis. As with other species, inclusion of diverse drivers of species distribution greatly improves the accuracy and usefulness of projections [29,30].

Further, the spatial [31] and temporal [27,32] scale of the data used to calibrate and test SDMs can influence results [33]. Discounting these influences obscures valuable information regarding the sensitivity of species to the influence of different bioclimatic conditions. Our findings are consistent with the notion that climatic and biological factors structure distributions at different operative scales [34] and suggest that using a multi-scale approach to developing SDMs is warranted. Our findings also lend support to the idea that evaluating the temporal transferability of models is important given the temporal dynamism of factors correlated with species distribution in disrupted systems and the sensitivity of associated projections [26]. Sensitivity to such issues in model construction are evident in forecasts for pandas relying on SDMs that have not included diverse drivers and non-climatic factors, which have yielded much more alarming results for panda population collapse than SDMs including some of these factors [16].

Our results show how climatic and non-climatic factors interact across spatial and temporal scales and suggest the species is amenable to management interventions to lessen the impacts of climate change. Good forest stewardship to reduce anthropogenic threats and increase forest cover, and, especially, preserve older, larger trees [9,11] may go a long way towards offsetting climate change impacts. Indeed, the three-decade period covered by our analysis witnessed temperature increases of 0.71°C accompanied first by a pattern of panda population decline, then population increase [18]. Thus, although climate influences panda distribution on the landscape, a warming climate did not prevent population expansion under good management practices, suggesting that the fate of the panda may not be as closely tied to climate as has been hypothesized previously [10,12–16]. Caution is warranted in generalizing our findings to the Qingling mountains, where pandas have been predicted to suffer the strongest negative effects from climate change in the future [10]. Further, because there are bamboo species that occupied the bioclimatic envelope at lower elevations that historically sustained panda populations, it may be feasible to target these species for large-scale assisted migration programmes to re-establish climatically suitable (locally endemic and/or natural occurring) bamboo in current panda habitat [35]. We suggest these and other management practices may allow the persistence of viable panda populations under climate change.

Climate change predictions present a bleak forecast for global biodiversity [3], and a profound challenge to wildlife managers tasked with developing recovery plans typically implemented on local and regional scales. The use of long-term, high-resolution historical data showing how species distributions have already responded under climate change marks an important step forward towards understanding bioclimatic envelopes, forecasting climate change effects and developing more successful climate-adaptation strategies. Anthropocene landscapes further complicate the application of SDMs to conservation because the current distribution of a species may no longer be an accurate representation of suitable habitat [36]. Inclusion of factors such as human land-use patterns through time can help reveal nuances in ecological factors supporting a species. It is essential that we correctly identify climate-mediated distributional shifts for the giant panda so that policies are not misguided and resources are not wasted. China has invested heavily in its protected area system, much of it to support panda recovery—and is reaching international targets to protect 17% of the terrestrial landscape [37,38]. This investment has already paid considerable dividends for panda recovery [9], and it is vitally important that it continues to do so under climate change. If climate modelling forecasts mischaracterize the potential impacts of climate change, incorrectly indicating that most of today's panda reserves will largely contain habitat unsuitable to pandas in the future, they may have unexpected adverse effects on decisions to continue investing in China's Protected Area system. While our analysis does not eliminate the possibility of reduced suitability of current reserves in the wake of climate change, it raises important considerations for management interventions that may limit climate change impacts on pandas.

Data accessibility

Datasets used in this study are available online from the Dryad Digital Repository: https://doi.org/10.5061/dryad.7m0cfxpr8 [39].

Authors' contributions

Z.Z. conceived the project, which was led by Z.Z., M.A.O. and R.R.S. J.T., X.Z., W.W., N.W.P. and M.A.O. conducted the analyses. J.T., R.R.S. and M.A.O. wrote the manuscript, with contributions from Z.Z., N.W.P. and F.W. Z.Z., X.Y., X.G., Z.Y., Q.D., J.Z., S.Y., M.H., H.Z. and H.H. conducted field surveys.

Competing interests

We declare we have no competing interests.

Funding

This research was supported by National Natural Science Foundation of China (31670530, 31600306, 31801992 and 31900337), Ministry of Science and Technology (grant no. 2016YFC0503200), Key projects of Education Department of Sichuan Province (18ZA0475) and the funds of China West Normal University (17E067, 17E068, 416446 and 416447).