Shifts in bird ranges and conservation priorities in China under climate change

50, 51]. As a reference, we distinguished the localities and mapped the breeding ranges of migratory bird species in China for comparison with other similar research results (for details, see S1 File & S2 Table). To forecast the response of species distribution to future climate change, it is necessary to select appropriate variables to construct species climate niches by considering the complex interactions among bioclimatic variables and digital elevation model (DEM) data [52–55]. To avoid overfitting caused by the autocorrelation between variables (especially in projections of future scenarios), we calculated the Pearson’s correlation at all presence sites for every pairwise combination of 19 bioclimatic variables downloaded from the WorldClim 1.4 database [56] (ESRI grids; bio 30s). Among all pairwise combinations, the variable with the most "high cor- relation pairs" (|r| > 0.9) was excluded [57]. Next, we selected the remaining 12 bioclimatic variables for modeling: bio 1–5 (Annual Mean Temperature, Mean Diurnal Range, Isotherm- ality, Temperature Seasonality, and Max Temperature of Warmest Month), bio 8–9 (Mean Temperature of Wettest Quarter and Mean Temperature of Driest Quarter), bio 12 (Annual Precipitation), bio 14–15 (Precipitation of Driest Month and Precipitation Seasonality), and bio 18–19 (Precipitation of Warmest Quarter and Precipitation of Coldest Quarter). To predict the ranges of birds in the future, we used three different General Circulation Models (GCM) for 2070 based on CMIP5 data, in both RCP 2.6 and RCP 8.5 scenarios: CCSM4, Had-

GEM2-ES, and MIROC5. These GCMs have been proven to perform well for vertebrates [6, 10, 21, 22]. We use the RCP2.6 and RCP 8.5 scenarios because they represent the lowest and highest level of emissions respectively. From this approach, we could derive the lower and upper limits of the impact on the bird species distribution [58]. All bioclimatic variables of these GCMs were downloaded from WorldClim 1.4. Data for National Nature Reserves (NNRs) in China were collected from existing research [36]. All grid data were rescaled to 30 arc-seconds (~1 kilometer) in ArcMap 10.2. 2.2 Distribution range modelling We used MaxEnt 3.4.1 to model range maps [42]; this software provides robust results by implementing a machine learning approach for presence-only data (such as citizen science data) [59, 60]. The model established based on cross-validation by five replicates (each repli- cate used 80% of the data for training and 20% of the data for test) for each bird with Cloglog output. Cloglog transformation is an updated algorithm in MaxEnt 3.4.1 for estimating the probability of presence by yielding a Bernoulli generalized linear model based on the Poisson distribution [61]. Compared with the traditional Logistic transform used in previous MaxEnt versions, Cloglog transformation can improve model performance to reduce the effects of sample selection bias while maintaining the same area under the receiver operating character- istic curve (AUC) value [42]. The background in MaxEnt was set as the entire land area of

China for all species. After modeling the distribution range in the current scenario, we pro- jected the model with different GCMs for RCP 2.6 and RCP 8.5 in 2070. We wrote R scripts to automatically batch the modeling process (for the code, see S2 File). The average Cloglog output for each bird and scenario was used to produce the species range map. The maximum test sensitivity plus specificity (MTSS) values for each species were used as thresholds to convert Cloglog raster outputs to “presence/absence” binary maps [62, 63], named predicted distribution range maps (DRMs). In order to improve the performance of the model, we conducted two rounds of simulation. In the first round, all 1,111 species with 161,630 localities and all 13 environment variables in current scenario (12 bioclimatic variables and the DEM) were applied to MaxEnt model by five cross-validated replicates with no special settings. We analyzed the results of first round PLOS ONE Bird range changes call for innovative conservation in China PLOS ONE | https://doi.org/10.1371/journal.pone.0240225 October 8, 2020 4 / 20