Introduction

Suitable habitats are pivotal for sustaining wildlife due to their capability of providing essential resources like food, shelter, and opportunities for reproduction (Kosterman et al. 2018). However, the rapid expansion of human activities has led to significant loss, degradation, and fragmentation in the natural habitats, which is considered one of the major threats to biodiversity (Ripple et al. 2014; Ceballos et al. 2015). Specifically, the majority of the world’s large carnivores, characterized by requirements for large space and high sensitivity to human pressure, have undergone substantial range contractions and population declines, with loss of suitable habitats as one of the main reasons (Ripple et al. 2014; Wolf and Ripple 2017). Despite the human modification on natural habitats, some large carnivores are able to sustain in human-dominated landscapes due to their behavioral plasticity and generalist nature, after lethal human activities (e.g., retaliatory killing following human-large carnivore conflicts, hunting and trapping) are ceased (Carter et al. 2012; Chapron et al. 2014; Lamb et al. 2020). Examples of coexistence between humans and large carnivores in Europe (e.g., brown bear Ursus arctos, wolf Canis lupus, Eurasian lynx Lynx lynx and wolverine Gulo gulo, Chapron et al. 2014), North America (e.g., brown bear, American black bear Ursus americanus, cougar Puma concolor and wolf, Penteriani et al. 2016) and South Asia (e.g., leopard Panthera pardus, Kshettry et al. 2017; Braczkowski et al. 2018), suggest the possibility of conserving endangered large carnivores in the modern human-dominated landscapes. However, fulfilling the habitat requirements of these apex predators in such environments remains a challenging task for wildlife managers and conservationists (Chapron et al. 2014; Wang et al. 2017).

Leopard, which is the most widespread felid, can utilize many different types of landscape (Jacobson et al. 2016), including those dominated by humans near urban areas (Kuhn 2014; Odden et al. 2014). The leopards in North China were previously recognized as a distinct subspecies endemic to China—the North Chinese leopard (P. p. japonensis). However, they have recently been reclassified and included in the Amur leopard (P. p. orientalis) category (Kitchener et al. 2017; Stein et al. 2023). They are categorized as Critically Endangered (CR) in the IUCN Red List (Stein et al. 2023) and Endangered (EN) in China Red List (Jiang et al. 2021). This big cat has undergone severe population decline primarily due to poaching, retaliatory killing, prey deficiency and habitat loss in the last century (Vitekere et al. 2020). The prohibition of wildlife hunting and poaching in China since 1990s has built a foundation for the recovery of this animal in the North China. Despite the extensive human activities, previous studies have demonstrated that leopard, as the only remaining large carnivore in the ecosystem, has been sustained at several sites in North China (Yang et al. 2020; Fu et al. 2023), suggesting the potential of restoring this endangered apex predator in the human-dominated landscapes across this vast area. Currently, the leopards in North China are considered to have multiple small regional populations (< 50 individuals) occurring in isolated habitat patches (Stein et al. 2023). Since 2014, the Chinese Felid Conservation Alliance (CFCA), a conservation NGO focusing on the investigation and protection of wild cats in China, and its collaborators have launched a project aimed at restoring the connectivity between the current local populations of this endangered big cat. However, the key environmental factors driving their distribution, the spatial pattern of their suitable habitats and the degree of landscape connectivity between the core habitats have rarely been studied (Cao et al. 2020; Liang et al. 2022, 2024; Fu et al. 2023), especially at the landscape level, which brings great challenge for their future conservation and restoration.

Here, based on occurrence data obtained from camera-trapping surveys and news reports of leopard sightings and encounters during 2014–2020, we constructed ensemble species distribution models to identify the most important environmental factors determining the habitat suitability of leopard in central North China. We mapped the spatial distribution of suitable habitats for the leopard. We further mapped the potential corridors at the landscape scale based on the least-cost pathway model, and identified the pivotal areas that may restricted the connectivity of these corridors based on the circuit theory. Together, our results offered new evidence and insights for assessing the current conservation status of this poorly studied big cat, and also built scientific foundations for further conservation and restoration actions towards its long-term persistence in the human-dominated landscape in North China.

Methods

Study area

We conducted this study in central North China (34°11´ ~ 43°49´ N, 103°11´ ~ 123°54´ E) (Fig. 1). With an area of about 936,000 km2, this region is bounded by the natural geographical barriers, i.e., the Yellow River, the Wei River, and the Gobi desert, that may restrict the dispersal of the leopards (Fig. 1). Specifically, this region encompasses the entire territory of Beijing, Tianjin, Hebei, and Shanxi provinces, as well as parts of Gansu, Shaanxi, Ningxia, Nei Menggu, Liaoning, Shandong, and Henan provinces (Fig. 1). The study area is dominated by a warm temperate semi-humid continental climate, with hot and rainy summers and cold and dry winters (Chen et al. 2021). The annual average temperature is about 8–13℃, and the annual precipitation ranges from 300 to 1,000 mm. The terrains in the study area are characterized by plains in its eastern part and mountainous areas in its northern and western parts, with elevation ranging from 0 to 3,800 m. The main vegetation of its original types is warm temperate deciduous broad-leaved forests and forest-steppes (Zhang et al. 2021). Leopard is the largest carnivore currently occurring in the study area, with pivotal ecological functions and potentials of being both an umbrella and a flagship species (Miththapala et al. 1996; Song et al. 2014).

Fig. 1
figure 1

The study area and the occurrence locations (N = 196) of North Chinese leopard in North China (2014–2020). The global human modification (GHM) index (map background) is from Kennedy et al. (2019)

Full size image

Data collection

Leopard occurrence data

The occurrence records of leopards were collected from two sources, camera-trapping surveys and public media reports, during 2014–2020. We identified 168 leopard occurrence locations with accurate coordinates from camera-trapping surveys conducted by the authors (e.g., Liu et al. 2020; Liu et al. 2023, N = 154) and from other peer-reviewed publications (e.g., Laguardia et al. 2015; Yang et al. 2020, N = 14) across the study area (Fig. 1). We also identified another 28 occurrence locations from public news reports, by searching Baidu (the largest Chinese search engine; www.baidu.com) and WeChat (the most popular Chinese mobile messaging app) for leopard sighting or encounter reports from officially accredited news outlets. We only used reports that contained both 1) the specific location of the leopard sighting or encounter, and 2) the photographs or videos of sighted leopards, to assure their reliability. We then obtained the coordinates of the sighting locations in Google Earth Pro (version 7.3.4) by first searching for the most detailed site names in the news reports, and then empirically determined the most likely occurrence locations of the leopard based on vegetation and terrain conditions. Together, the preliminary dataset used in the study included 196 leopard occurrence locations.

To avoid over-fitting of spatial models in areas with clumped occurrence points (Dormann et al. 2007; Kramer-Schadt et al. 2013), we conducted a spatial filtering procedure for all locations prior to model construction by using the spThin package (Aiello-Lammens et al. 2015) in R environment v. 4.1.1 (R Core Team 2021). The occurrence locations were filtered by randomly selecting one from all locations within the same 1 km × 1 km grid cell and eliminating the rest, if any, resulting in 155 sites left for habitat modeling. In addition to the occurrence locations, we obtained a random sample of 10,000 background locations (i.e., the pseudo-absence locations) within the extent of the study area excepting occurrence cells for model fitting. The 155 occurrence locations (presence data) and the background locations (pseudo-absence data) obtained in this step were directly used as the input dataset in both the subsequent construction of the univariate and multivariate ensemble models.

Habitat variables

We incorporated four groups of variables that have been reported to impact the habitat suitability of leopards (Yang et al. 2020; Zhu et al. 2021; Liu et al. 2023; Liang et al. 2024), including variables measuring the topographic characteristics, climate conditions, vegetation, and anthropogenic pressures (Table S1). The topographical variables included elevation (DEM) and surface ruggedness (RUG). The RUG was generated by the digital elevation map (30 m × 30 m) with the terrain ruggedness index tools in ArcMap 10.5 (ESRI, Redlands, CA, USA) using a moving window of default 3 × 3 grid cells. The climatic variables included annual mean temperature (TEM) and annual precipitation (PRE) (Fick and Hijmans 2017). The vegetation variable included percent tree cover (TC) and shrub and grassland cover (SGC). The SGC was derived from the GISD30 land cover dataset (Zhang et al. 2022), which included grid cells identified as herbaceous cover, shrubland, evergreen shrubland, deciduous shrubland and grassland. We used global human modification (GHM) (Kennedy et al. 2019) and road density (RD) as proxies for the intensity of anthropogenic pressures. Previous studies have shown that road networks have a significant negative impact on large mammals (Zeller et al. 2020; Blackburn et al. 2021). Although GHM is a comprehensive layer that integrates five major categories of human impacts, the transportation data used to calculate the index is derived from OpenStreetMap (www.openstreetmap.org), which has poor coverage for China. Therefore, we decided to use RD as a supplement to GHM in the study area. The vector dataset of road networks was collected from the 2016 Navigation Map, which covers both the paved roads (highways and other roads at various level from national to local) and railways in China (Cao et al. 2020). We calculated the length of all types of roads mentioned above within each 1 km × 1 km grid cell as the road density.

Given that different environmental predictors may be related to habitat selection of species at different spatial scales (Wiens 2001), multiscale optimization is considered an effective approach to determine the appropriate scale for each individual variables (McGarigal et al. 2016). Based on the home range of the leopard, we selected four candidate spatial scales (including 500 m, 1,500 m, 3,000 m and 6,000 m) for each variable. We set the upper limit to 6,000 m to approximate the average home range of leopards (~ 114 km2) reported in previous studies from similar ecosystems to our study area (Broekman et al. 2022). For each grid cell in the original variable layer, we re-calculated its’ value as the mean value of all grid cells within a certain radius (500/1,500/3,000/6,000 m) centered on it. The calculations were conducted using the Focal Statistics tool in ArcMap 10.5 (ESRI, Redlands, CA, USA). Finally, all variable layers were resampled to a UTM projection, with 1-km cell size.

To avoid multi-collinearity, we calculated the Pearson correlation index for each pair of variables and calculated the variance inflation factor (VIF) for all variables (Zuur et al. 2009). Finally, we retained DEM, RUG, PRE, TC, SGC, GHM and RD for subsequent analysis, with correlation coefficients between any pair of them less than 0.7 and VIF less than 3. We excluded one variable (i.e., the annual temperature) because it was correlated to the elevation (correlation coefficient = -0.72).

Habitat suitability modeling

Variable preselection

We conducted a univariate scaling for each variable (Mateo-Sánchez et al. 2015; Atzeni et al. 2020) to identify the scales most strongly related to leopard presence, by using an “ensemble” SDM with the biomod2 package (Thuiller et al. 2009) in R environment v. 4.1.1 (R Core Team 2021). Taking advantage of combining results from multiple modeling algorithms, this model is proven more robust than single algorithm and has been increasingly used in wildlife studies (Ahmadi et al. 2017; Almasieh and Cheraghi 2022; Almasieh et al. 2022).

The ensemble model was built from the highest performing models (AUC ≥ 0.80) of six different widely used algorithms (including three regression-based methods, i.e., generalized linear model [GLM], generalized additive model [GAM] and multivariate adaptive regression splines[MARS], and four machine learning algorithms, i.e., generalized boosting model [GBM], random forest [RF], artificial neural network [ANN], and maximum entropy [MaxEnt]). We fitted the models with randomly selected 80% of the occurrence locations and all pseudo-absences as training data and evaluated the models’ performance using the rest 20% of the occurrence points as the test data. We ran each individual model 100 times. The ensemble model was obtained by averaging the individual models. The performances of the univariate ensemble models across scales were evaluated by area under the curve (AUC) of the receiver operating characteristic curve (ROC) and the true skill statistic (TSS) (Allouche et al. 2006). For each variable, we chose the scale with the highest AUC value as the optimal spatial scale for that variable (Table S2).

Multivariate model building

We built the final multivariate ensemble model using a suite of scale-optimized variables obtained above. We estimated the relative importance of each predictor using the biomod2 package by calculating the Pearson correlations between the standard predictions (i.e., fitted values) and predictions where the variable under investigation has been randomly permutated (Thuiller et al. 2009). A higher correlation indicates less difference between the two predictions and less importance for the permutated variable. The final ensemble model was built from the highest-performing models (AUC ≥ 0.90) of seven different algorithms (i.e., GLM, GAM, MARS, GBM, RF, ANN, and MaxEnt). The ensemble model was obtained by averaging the individual models, to produce the habitat suitability map with values ranging from 0 to 1, where higher values represent higher habitat suitability. We also reported the response curves based on the final ensemble model to explore the correlations between the habitat variables and the leopard habitat suitability.

Based on the final distribution model, we identified the most suitable habitats as the core habitat patches of leopards. We first converted the continuous predicted distribution probability into a binary classification of suitable/unsuitable map using a modified lowest-presence threshold (LPT; Pearson et al. 2007; Waltari et al. 2007; Waltari and Guralnick. 2009). We further defined suitable patches ≥ 100 km2 as the core habitats for subsequent analysis, where this threshold was chosen as approximately the average home range size of leopards in ecosystems similar to our study area (Stein et al. 2023; Broekman et al. 2022). The threshold for LPT was set to classify 90% of all occurrence locations in the training data as the suitable habitat, or a 10% omission rate, abbreviated as LPT90%. Compared with the most commonly used LPT that yielded zero omission error, LPT90% was more conservative and identified smaller areas, thus resulting in more restricted pictures of potential distributions of the target species (Waltari and Guralnick 2009). To assess the protection status of the leopards’ habitat, we also calculated the overlapped areas between current nature reserves and the model predicted suitable habitats. To identify our knowledge gap on the occurrence status of leopards in each habitat patch, which may guide our future survey and monitoring efforts, we further classified the core habitats into two categories, i.e., leopard presence confirmed and unknown, based on the on-site field survey data (Liu et al. 2020; Wan et al. 2020; Fu et al. 2023).

Landscape connectivity analysis

Species are usually more tolerant of landscapes in movement corridors compared with the core habitats (Beier et al. 2008). Based on the analysis of actual species movement data, several studies have shown a negative exponential relationship between the resistance value of heterogeneous landscapes to species movement and habitat suitability, rather than a simple negative linear relationship (Trainor et al. 2013; Mateo-Sánchez et al. 2015; Keeley et al. 2016). We obtained the resistance layer based on the continuous values of model predicted habitat suitability using the following equation (Keeley et al. 2016):

$$R=100-99\times \frac{1-{e}^{-c\times h}}{1-{e}^{-c}}$$

where R is the resistance value, h is the habitat suitability predicted by the ensemble model, and c is a parameter that determines the shape of the negative exponential curve. We used 4 as the value of c, which is considered to be realistic in previous study of leopards in this region (Cao et al. 2020).

The potential dispersal corridors between core habitats were identified based on the minimum resistance model using the Linkage Pathways Tool in Linkage Mapper (v. 3.0) (McRae and Kavanagh 2011). This program identifies adjacent core areas, creates a network of core areas using adjacency and distance data (i.e., Euclidean distance, EucD), calculates cost-weighted distances (CWD) and least-cost pathways (LCPs), and combines least-cost corridors into a single map (Dutta et al. 2016; Bu et al. 2020). We set the maximum EucD for a corridor as 82 km according to the longest dispersal distance recorded for a leopard (Farhadinia et al. 2018). The LCP is the single path associated with the minimum cost-weighted distance between a source and destination (Dutta et al. 2016; Cerreta et al. 2023). Two metrics were calculated to quantify the quality of each linkage between core habitats identified: CWD: EucD and CWD: LCP ratio. Higher CWD: EucD or CWD: LCP ratio between two core habitats indicates higher degree of difficulty to move between them.

To further assess the conservation and restoration priorities, we further identified the pivotal areas within the LCPs which may restricted the connectivity of them, which are known as the “pinch points”. We used the Pinchpoint Mapper Tool in Linkage Mapper, interfacing with Circuitscape (v. 4.0.7, McRae et al. 2013). This approach, based on the circuit theory, effectively identifies narrow sections within a corridor, known as pinch points, where animals have less alternative selection in their movements, but have to use the pinch points. Therefore, the loss of pinch points can disproportionately disrupt connectivity of the corridors, making them important focal areas for conservation and restoration efforts. Additionally, corridor effective resistance values were determined, serving as a measure of connectivity that complements least-cost distances. We set the cost-weighted width cutoff value as 33 km according to the average dispersal distance of leopards (Farhadinia et al. 2018). The analysis was conducted in “All to one” mode, allowing current flow from all source nodes (i.e., core habitat patches) to iteratively reach each ground node to generate cumulative current density maps. Areas with high current densities (i.e., fewer alternative pathways) were identified as the key pinch points, which might be crucial for maintaining connectivity for the entire network (McRae et al. 2008). According to Pelletier et al. (2014), there are no clear guidelines on systematically identifying pinch points or using a single current density value threshold across multiple tiles. Instead, we identified pinch points by considering the darker (i.e., purpler in our result) areas on the map produced by Circuitscape.

Results

Habitat suitability

The ensemble models showed adequate predictive performance according to the AUC and TSS metrics (AUC = 0.996, TSS = 0.931) (Table S4). The prediction accuracy was adequate for all individual algorithms (TSS > 0.8) (Table S4), where the RF algorithm had the highest AUC value (Table S4).

The predicted map indicated that, the current suitable habitats of leopards are highly fragmented, with large patches mainly located in Shanxi, Shaanxi, and along the border between Gansu and Ningxia provinces (Fig. 2). Among all the 8,679 km2 of suitable habitats, only 23.32% (2,024 km2) were covered by current nature reserves (Table 1). We identified 14 core habitats (139–1,084 km2, mean = 495.21 km2) with a total area of 6,933 km2 (Table 1; see more details in Table S6), among which only 25,26% (1,751 km2) of area in 11 core habitats were covered by current reserves (including 3 provincial-level and 9 national-level nature reserves, as 1–2 reserves in each patch. Table 1; Table S6). The other 3 core habitat patches were not covered by any nature reserves. The on-site field surveys showed that, 11 core habitats were confirmed with leopard occurrence, and the other 3 core habitats were undetermined for the presence of leopards (Table S6).

Fig. 2
figure 2

Model predicted habitat suitability of leopard in central North China

Full size image
Table 1 Area and proportion of suitable/unsuitable habitats of leopard in central North China
Full size table

The mean relative importance showed that DEM, PRE and TC were the main factors affecting the habitat suitability of leopards (Table S5). The response curves showed that the leopard preferred areas with high elevation (> 1,500 m) and high annual precipitation (> 600 mm) (Fig. 3). Leopard habitat suitability and tree cover were positively correlated when tree cover was below 20%, and the suitability was generally high when tree cover was above 20%, but with a dip in the 30–50% range (Fig. 3). RD and GHM both showed a unimodal pattern, with a low to medium levels being preferred (Fig. 3; Table S5).

Fig. 3
figure 3

Response curves of habitat suitability to predicting variables based on the ensemble model. DEM: elevation (m); RUG: ruggedness; TC: percent tree cover; GHM: global human modification; RD: road density (km/km2); SGC: percent shrub and grassland cover; PRE: annual precipitation

Full size image

Landscape connectivity

The results of landscape analysis indicated that, the connectivity between core habitats were highly limited (Fig. 4a). C10, C11, C13 were isolated, as there was no LCPs connecting these patches to others (Fig. 4a). Other core habitat patches showed clustering, mainly in three regions: the border between Gansu and Ningxia provinces, the region of Shaanxi and Shanxi provinces, and the border between Hebei and Shanxi provinces (Fig. 4a, b). The mean Euclidean distance between all patches was 40.06 km (range: 1.00–74.09 km, Table S7). The model identified 8 LCPs among the 14 core habitats with an average length of 57.22 km (range: 1.41–134.98 km, Table S7). L4, L6, and L7 had both the highest CWD: LCP ratio and CWD: EucD ratio, indicating the highest cost for moving along them (Table S7). Our cumulative current density map highlighted the existence of numerous pinch points across the potential corridors (Fig. 4b). The linkages involved in pinch points included L3, L6and L7, which could be pivotal corridors for leopard dispersal (Fig. 4b).

Fig. 4
figure 4

Habitat connectivity between core habitats (a) and pinch points identified across the potential corridors (b). LCPs: least-cost pathways. Black numbers in Fig. 4a and Fig. 4b indicate the ID of the potential corridors and core habitats, respectively. The insert map in Fig. 4a is enlarged as Fig. 4b to show the details of the pinch points across the potential corridors. The base map in Fig. 4b is from ESRI, where the light green patches indicate continuous forests

Full size image

Discussion

Key factors determining leopard habitat suitability

Leopards, as a widely distributed species, are capable of utilizing various habitats, including tropical and temperate forests, Savanna, grasslands, and even deserts (Jacobson et al. 2016; Stein et al. 2023). Leopards can also persist near major towns, such as Mumbai (Odden et al. 2014) and Johannesburg (Kuhn 2014). In this study, our results indicated that, the leopards in North China prefer higher (elevation > 1,000 m) and wetter (annual precipitation > 600 mm) forests, with negative response to shrublands and grasslands. They also showed moderate avoidance to human disturbance, but the impacts of human modification index and road density were less important compared to natural habitat characteristics. These results corroborate the findings of Yang et al. (2020) and Liang et al. (2024). Specifically, within our study area, we observed that the mountainous regions (the western part of our study area) also possess a lower human modification index than the lowlands (the eastern part of our study area; see Fig. 1 and Supplementary Fig. 1). Given the great difference between these two regions, constructing separate model for each region may theoretically improve the overall predicting performance. However, we were unable to do so due to lack of surveys and data in the eastern region with high intensity of human disturbance. This might bring difficulties in identifying the true ecological drivers for leopard habitat suitability (especially between elevation and human modification on the landscapes), and might also neglected the potential different habitat requirements for the leopards in these two regions. Therefore, we recommend that future leopard surveys should maintain the current survey network while also covering more lowland areas in the eastern region of North China.

In contrast to our findings about leopards’ strong preference for moist forests with annual precipitation higher than 600 mm, many studies have demonstrated leopards’ ability to inhabit semi-arid areas with less vegetation (e.g., Shahsavarzadeh et al. 2023), and can even utilize the desert areas (Stein et al. 2023). There are several potential explanations for this difference. First, the leopards in the study area may have specific habitat requirements different from other populations, as the leopards in China have rarely been recorded in the semi-arid, arid and alpine areas since 1950s (according to the occurrence records of leopard in A Guide to the Mammals of China, Smith et al. 2009; Zhu 2022). Second, previous studies have demonstrated the importance of habitat structures that can provide good cover and are less permeable to people as refugia for large carnivores inhabiting human-dominated landscapes (Oriol-Cotterill et al. 2015). Because of the long history of human modification on the natural ecosystems in North China, the leopard’s preference for forests may be also related to the role of forests in mitigating human effects, especially given the evidence that leopard population can decline drastically when human disturbance exceed certain threshold (Woodroffe 2000). Despite these uncertainties in leopard’s preference for humid forests, we emphasize the importance of the recovering forests for the current habitat suitability and their future restoration in North China. The Chinese government has launched ambitious programs in restoring the forests in the past decades (Chen et al. 2019), resulting in a remarkable net increase in forest cover in North China since 1990 (Cheng et al. 2024), which may be one of the crucial elements for the future survival of the endangered leopards.

Besides the environmental characteristics included in our habitat modeling, the prey availability is also an essential factor that may affect the habitat suitability of big cats (Karanth et al. 2004; Ahmadi et al. 2017; Wang et al. 2017; Searle et al. 2020). As a generalist, leopards can utilize a broad range of prey from large herbivores, mesocarnivores (e.g., badgers, foxes) to hares and pheasants, whereas they have a preference for medium-sized ungulates (Hayward et al. 2006). However, current research on the dietary habits of leopards in North China is highly limited (Wang 2020). Meanwhile, although there are several potential ungulate prey species occurring in our study area, including Siberian roe deer (Capreolus pygargus), Chinese goral (Naemorhedus griseus), wild boar (Sus scrofa) and free-ranging livestock (cattle, sheep, goat), their abundance have not been systematically evaluated and mapped in previous studies. Thus, we are currently unable to incorporate prey availability as a predictor into our model due to lack of knowledge and data on the leopards’ dietary composition and prey abundance. Recent advances in quantitative diet analysis using non-invasive fecal samples and DNA meta-barcoding provide us a powerful and efficient tool to determine the prey components of carnivore species (Shao et al. 2021). And, the rapidly increasing development of local and regional camera-trapping networks have a good potential of generating robust estimates of relative abundance of the major prey species across the region (Li 2020). Future studies shall further examine the leopard’s diets and its seasonal and spatial variations in North China, and produce maps of prey distribution and abundance using data from regional monitoring networks. Incorporating these factors into future predictions will greatly improve the model power and robustness, which will also provide us valuable insights to guide future conservation and recovery of large carnivores in North China.

Habitat connectivity and conservation implications

The 14 core habitats of leopards we identified are highly fragmented. Besides, the occurrence of leopards is still not clear in 3 of these identified core habitats due to lack of field surveys. Among the patches with confirmed leopard records, a few centered around the nature reserves are under long-term camera-trapping monitoring, such as patch C4 of 819 km2 (Tieqiaoshan and Bafuling Provincial Nature Reserves in Shanxi province; Liu et al. 2020) where 56 individuals have been identified (unpublished data) and patch C13 of 963 km2 (Ziwuling National Nature Reserve in Shaanxi province where 27 individuals have been identified; Yang et al. 2020). Although these patches are larger than the core area of individual leopard’s home range reported in previous studies (13.5 to 80.6 km2; Fattebert et al. 2016; Ray-Brambach et al. 2018; Roex et al. 2021), the demographic dynamic of leopards in these patches have received little attention, which brings great uncertainty of the future sustainability of the leopard populations within these areas. For the patches where leopard presence has not yet been confirmed (i.e., C1, C8 and C14), we suggested intensive investigations to determine the current occupancy status of leopards. For example, patch C8 is located at a key junction connecting the core habitats in Shanxi and Shaanxi provinces, with an area of 193 km2, which may be able to serve as a stepping stone for leopard. Our results of the connectivity analysis identified potential paths for leopard movement between different habitat patches. However, the actual usage of these paths for leopard is still poorly understood, especially for those with evident pinch points. Landscapes that retain more connections between patches of isolated habitats can maintain dispersal pathways for large mammals and increase their demographic and genetic population size (Mills and Allendorf 1996). The persistence of large carnivores in small habitat patches amid the human-dominated landscapes highly rely on the immigration of individuals from large patches to surrounding small patches, which will serve as the source and sink, respectively, in a meta-population (Hanski 1998; Perry and Lee 2019; Lamb et al. 2020; Bertassello et al. 2021). Thereby, we suggest that the future conservation of leopard in North China shall be based on systematically designed monitoring of all local populations of the leopards across the region, with emphases on the dispersal and migration of individuals between those patches. Besides, in our analysis of landscape connectivity, we included dispersal distance thresholds to approximate potential leopard dispersion among patches. However, occurrence records themselves can not represent the functional connectivity (i.e., how animals actually move through the landscape) (Crooks and Sanjayan 2006; Vogt et al. 2009; Rezvani et al. 2024). Therefore, we further suggest incorporation of telemetry studies to track a sample of leopards using satellite collars, which could significantly enhance our understanding of the functional connectivity and the efficiency of predicted corridors. Future conservation efforts shall also focus on the restoration of landscape corridors given their importance of maintaining the potential migration between different patches. Despite the obvious gap between the suitable habitats and current nature reserve coverage, establishing more reserves to fully cover all core habitats is not practicable given the vast area, heavy land use and the livelihood of local communities within the region. Such a situation further emphasizes the importance of achieving coexistence between large carnivores and humans through alternative approaches, such as the OECMs (Other Effective Area-based Conservation Measures; Dudley et al. (2018)), especially given the ability of leopards to adapt multiple disturbances (Liu et al. 2023). Together, the further monitoring and conservation of leopards in North China may be able to serve as an example for sustaining the endangered mega-mammals in the human-dominated landscapes in developing countries.