4.1. Study area
The Sanjiangyuan National Nature Reserve (SNRC) spans from 89˚45´E to 102˚23´E and 31˚39´N to 36˚12´N, located in southern Qinghai Province, China, and on the central-eastern Tibetan Plateau (Fig. 5). It forms the source of the Yangtze, Yellow, and Mekong Rivers, earning it the title “Water Tower of Asia”. Encompassing Golmud City and four Tibetan autonomous prefectures-Guoluo, Yushu, Hainan, and Haixi-the SNRC covers an area of 363,000 km², representing 50.4% of Qinghai Province’s total area. The landscape is characterized by a mix of plateaus and mountains, including the Tanggula, Bayan Har, Kunlun, and A’nyêmaqên ranges, with elevations ranging from 1,956 to 6,824 meters and an average elevation of around 4,500 meters45. The rate of precipitation variation from 1960 to 2015 was 6.653 mm/10a. The region experiences a plateau continental climate, marked by a cold season dominated by the Tibetan Plateau High Pressure and a warm season influenced by the Indian Ocean monsoon, contributing to its distinct environmental diversity6, 46. The natural environment exhibits considerable diversity, spanning from coniferous forests and shrubs to alpine meadows, grasslands, and sparse vegetation across a southeast to northwest gradient. The area is home to nationally protected wildlife, including the snow leopard (Panthera uncia), Przewalski's gazelle (Procapra przewalskii), wild yak (Bos mutus), Tibetan sand fox (Vulpes ferrilata), Tibetan wild ass (Equus kiang), and the Tibetan antelope (Procapra picticaudata).
4.2. Species distribution data collection
For our study, we deployed 55 infrared cameras (Ltl-C180; Beijing Dingxing Technology Co., Ltd., Beijing, China) in seven typical canyons within the SNRC from April to September 2023 (Fig. 5). Cameras were strategically placed over 500 m apart to ensure spatial independence and mounted on trees approximately 0.5 m above ground. Cameras were programmed with moderate sensitive sensor setting, to shoot 3 photos and a 20s video when being triggered, and time was set to 24 h per day. Cameras were maintained in the field for 4 to 6 months, and were inspected for SD cards and batteries upon movement of cameras. No bait was used to attract animals, which is important in situations where the aim of the study is to look at animal behaviors in an unbiased way.
Photographs and videos were summarized by sites, hour, and date at each camera placement site. To ensure independence of photographic capture events, we defined detection at a sample point as one individual photograph of one species during a 30-min period. The number of effective camera trap days was calculated as the time frame between camera setting, and the date of the last photograph or video was taken if a malfunction occurred (based on date stamp). Photographic Rate (PR) was used to compare independent detections in different wild ungulate species. Photographic Rate for each species in each camera site was calculated as the number of independent detections for every species divided by the total sampling effort for that sample point (number of camera-days), multiplied by 100:
PR = (No. of detections/Camera-days) * 100
4.3. Collection and filtering of environment variables
Models necessitate two distinct datasets: the geographic coordinates of the target species and a suite of environmental variables encompassing climate, terrain, land cover, and human interference factors (Supplementary Table 1). Climate data, representing averages from 1970–2000, were sourced for 19 bioclimatic factors via GIS47. Digital Elevation Model (DEM) information was obtained from NASA’s Alaska Satellite Facility (ASF), with slope and aspect data derived from 12.5m resolution DEMs. Variables indicating human interference were drawn from the 2017 National Catalogue Service for Geographic Information, including road and residential density and proximity to water sources, calculated via Euclidean distance. Land cover data were sourced from the European Space Agency (ESA) 2021 WorldCover dataset, categorized into 11 land use types. In GIS 10.6, through mask extraction and resampling operations, the environmental data of the study area were obtained for subsequent model construction.
The autocorrelations and multiple linear duplications among environment variables might affect the prediction results of the model48. To reduce the overlap of information between variables, we used the Band Collection Statistics tool in ArcGIS 10.5 to test the spatial correlations of the above environmental variables. The variables with high correlation (|r| ≥ 0.80) were eliminated, and those with low correlation and more biological implications were introduced into the model operation49, 50. Spatial resolution of the data was standardized to 30 m, with coordinates projected to WGS1984 UTM Zone 47 N and converted to ASCII format for model integration.
4.4. Habitat suitability model
The MaxEnt model was developed by Steven Phillips as a density estimation and species distribution prediction model based on maximum entropy theory51. MaxEnt requires the current geographical distribution of the species and its environmental constraints to explore the possible distribution of maximum entropy under this constraint The probability distribution of the species with the highest entropy is most similar to the current distribution of target52. When modeling habitat suitability, the MaxEnt model only requires presence data of the focal species, without the need for its absence data that are difficult to obtain in practice53. We used the MaxEnt software (v. 3.4.3) to model the habitat suitability for Alpine musk deer and Blue sheep, respectively54. As a popular niche model, the MaxEnt is often used for habitat evaluation due to its high accuracy, high computational efficiency, and ease of use.
We used the Jackknife method and the response curve to test the importance and the effect interval of the environmental variables in the model. To ensure model stability, we set the random test percentage to 25% and performed ten repetitions of the model by using the bootstrap method55. The regularization multiplier was set to 1, and the maximum number of background points was set to 10,000. After ten repetitions, the average habitat suitability index was used as the final result56.
The MaxEnt model automatically generates receiver operating characteristic (ROC) curves. This curve uses the false positive rate (1-specificity) as abscissa and the true positive rate (sensitivity) as ordinate. The area under the curve (AUC) values ranges from 0 to 1, with values closer to 1 indicating more accurate results57. If the AUC value is 0.5–0.7, the model evaluation reliability is low, if the AUC value is 0.7–0.9, the model evaluation reliability is medium, and if the AUC value exceeds 0.9, the model evaluation reliability is high58. We used the area under the receiver operating characteristic curve to verify the accuracy of the MaxEnt model output.
MaxEnt computed the contribution percentage of each environmental variable, and variables with relatively high contribution rate were selected to analyze the corresponding response curves. At the same time, a frequency distribution histogram was used to statistically analyze the relatively high environmental variables. The MaxEnt model was operated to obtain the predicted distribution map of the Alpine musk deer and Blue sheep, which was divided into four different suitable habitats. The result map was generated by ArcGIS.