Mitigating Human–Nature Tensions Through Adaptive Zoning Informed by the Habitat Suitability of Flagship Species: Insights from the Longbao Reserve on the Qinghai–Tibet Plateau

Abstract

Zoning is vital for balancing biodiversity conservation and sustainable development in protected areas, yet traditional approaches often lead to ecological overprotection and social conflict. This study introduces an integrative modeling framework to optimize zoning strategies in the Longbao Reserve on the Qinghai–Tibet Plateau. We employed MaxEnt and Random Forest algorithms to evaluate habitat suitability for two flagship species: the bar-headed goose (Anser indicus) and the black-necked crane (Grus nigricollis). Results showed that 7.9% of the reserve comprised highly suitable habitats, mainly in the southeast, characterized by wetlands, water proximity, and low human disturbance. Land use and June NDVI emerged as key predictors, contributing over 30% and 35% to model performance, respectively. Based on habitat suitability and current zoning mismatches, we propose a revised four-tier zoning scheme: Core Habitat Conservation (16.9%), Ecological Rehabilitation (7.2%), Ecological Management (53.5%), and Sustainable Utilization Zones (22.4%). This refined framework aligns conservation priorities with local development needs and offers a scalable approach to adaptive protected area management.

1. Introduction

Protected areas play a central role in ecological governance, underpinning efforts to curb biodiversity loss and address climate instability [1,2]. Functional zoning systems with tiered protection levels are essential for balancing conservation objectives with sustainable land use [3,4]. As China transitions toward a national park–centered protected area system, the development of adaptive, science-based zoning frameworks becomes critical for achieving both biodiversity conservation and governance goals [5,6,7].

Despite the global expansion of protected areas [8], many continue to suffer from rudimentary zoning structures, vague management mandates, and limited administrative effectiveness [9,10]. Reactive approaches have often resulted in poorly defined objectives, overly simplified spatial zoning, and weak ecological connectivity [11,12]. The commonly adopted “core–buffer–experimental” zoning model has shown limitations in responding to complex and dynamic ecological processes [13,14]. A recent study of Sanjiangyuan National Park by Ma et al. (2022) revealed that many biodiversity-rich areas fall outside current management zones, highlighting significant misalignments between spatial zoning and ecological priorities [15]. Similarly, Wang et al. (2023) reported that most national parks in China lack functional sub-zoning systems capable of adapting to evolving ecological conditions and increasing anthropogenic pressures [16]. These mismatches largely stem from ecologically arbitrary zoning boundaries. When species distributions are ignored, critical habitats risk being excluded from protection. Consequently, there is growing interest in ecological indicators such as habitat suitability. For example, Murillo et al. (2024) employed hierarchical models to identify priority areas for marine conservation, demonstrating how data-driven approaches can inform more effective spatial planning [17]. However, such methods remain underutilized in terrestrial contexts—especially within ecologically complex, high-altitude landscapes like the Qinghai–Tibet Plateau. This gap underscores the need for adaptive zoning strategies grounded in habitat suitability modeling—an approach this study aims to develop and apply.

In response, China’s strategy is shifting from quantitative expansion to qualitative enhancement. Policies such as the ecological corridor initiative and the national park system support landscape-scale integration [6,18]; however, tools for precision spatial zoning remain insufficient [13,16]. When effectively implemented, functional zoning serves as a spatial regulatory mechanism to reduce anthropogenic disturbance and buffer ecosystem vulnerabilities. By prioritizing zones based on ecosystem services, it enhances system-level stability and fosters ecological resilience [15,19]. Although recent studies increasingly advocate multi-layered zoning frameworks informed by ecological sensitivity and human pressure gradients [20,21], their application remains constrained by data limitations and methodological gaps [22,23]. In light of these challenges, scholars have called for spatial zoning strategies that integrate dynamic regulation and scenario-based planning. For instance, Li et al. [18] analyzed ecological risks in the Three-River Source Region and emphasized the need for flexible zoning that reflects landscape dynamics. Wang et al. [24] demonstrated that context-specific functional zoning can reconcile ecological protection with development objectives through adaptive mechanisms. Nonetheless, such frameworks remain underutilized, particularly in ecologically complex regions. In the Longbao Reserve, ambiguous boundaries between wetlands and grasslands have resulted in overlapping land use and fragmented management. This spatial ambiguity has hindered targeted conservation efforts and increased disturbance risks to key species during critical life stages. These site-specific issues underscore that zoning approaches lacking species-level resolution may fail to safeguard priority habitats or may impose overly generalized regulations. This highlights a broader need for ecologically informed, adaptive zoning strategies tailored to dynamic ecosystem conditions.

This study focuses on two flagship bird species in the reserve: the black-necked crane and the bar-headed goose. The black-necked crane is listed as Vulnerable by the IUCN [25] and receives Class I national protection in China [26]. It breeds in alpine wetlands above 3000 m across the Qinghai–Tibet Plateau and conducts short-range seasonal migrations within the region. The bar-headed goose, though not globally threatened, is a dominant high-altitude migratory species across Central and South Asia and exhibits strong site fidelity to wetland breeding grounds on the plateau [27]. To bridge the gap between ecological value and current zoning, we employ Maximum Entropy (MaxEnt) and Random Forest algorithms to model habitat suitability and identify key environmental drivers, including anthropogenic factors. Based on the habitat preferences of these two species, we propose a differentiated functional zoning framework aligned with ecological spatial patterns. The resulting regulatory system seeks to harmonize biodiversity conservation with sustainable land use, offering guidance for zoning refinement in heterogeneous wetland–grassland ecosystems. This study integrates species-specific modeling with spatial planning, providing a data-driven, ecologically grounded approach to adaptive management.

2. Materials and Methods

2.1. Study Area

The Longbao Reserve is located in Longbao Town, Yushu City, Qinghai Province, positioned within the central basin of the eastern Qinghai–Tibet Plateau transition zone (33°08′–33°14′ N, 96°25′–96°37′ E), at an average elevation of approximately 4200 m (Figure 1) [28]. The reserve spans 10,000 hectares, of which 3349 hectares are classified as wetlands, including 430 hectares of lakes, 1660 hectares of marshes, and 1050 hectares of marsh grasslands [29]. Established in 1984 and elevated to national status in 1986, the reserve was the first in China specifically designated to protect the breeding habitat of the black-necked crane. Owing to its role in conserving flagship endangered species, representative high-elevation wetland ecosystems, and critical water regulation functions, Longbao Wetland was designated a Wetland of International Importance in 2022 [30]. A total of 149 vertebrate species have been recorded within the reserve, including 132 species of birds. Flagship species include the black-necked crane (175 individuals recorded in 2023) and the bar-headed goose (5127 individuals recorded in 2023, accounting for approximately 9.4% of the global population). The region experiences a plateau continental climate, with a mean annual temperature of approximately 2.9 °C. Vegetation is dominated by alpine and marsh meadows, while land use is dominated by natural wetlands and pastoral grasslands [31].

The Tibetan Plateau functions as a vital ecological security barrier at the global scale, playing key roles in climate regulation, water resource provision, and biodiversity conservation [12,20]. Longbao Wetland, a representative high-elevation marsh ecosystem located on the eastern fringe of the Tibetan Plateau, serves as a critical protected area within China’s plateau wetland system. However, like many other nature reserves, the current zoning scheme is hindered by overly broad classifications, ambiguous boundary definitions, and difficulties in practical implementation [10]. The breeding and foraging habitats of key protected species—such as the black-necked crane and the bar-headed goose—partially overlap with buffer and experimental zones, which are insufficiently regulated. Persistent disturbances from livestock grazing and human encroachment arise from poorly defined zone boundaries. In particular, during the breeding season, livestock intrusions into wetland margins—coupled with artificial fencing—impede species movement, heighten disturbance risks, and intensify human–wildlife conflicts. Despite these ongoing management challenges, the reserve offers strong ecological and structural foundations for implementing scientifically informed zoning. Its moderate spatial extent, enclosed topographic structure, and distinctly stratified ecosystems—ranging from lakes to marshes to alpine meadows—provide clear ecological boundaries that are conducive to coherent spatial planning. The spatially clustered and temporally stable distributions of flagship species such as the black-necked crane and the bar-headed goose, combined with their predictable behavioral patterns, support the development of species-activity-based dynamic zoning strategies.

2.2. Research Data

To support habitat suitability modeling for flagship species in the Longbao Reserve, this study collected and integrated two key types of datasets: species occurrence data and environmental variables. Field surveys were conducted to monitor the breeding and foraging activity of the black-necked crane and bar-headed goose. Simultaneously, multi-source remote sensing and spatial datasets were obtained to represent land cover characteristics, vegetation dynamics, and human and natural environmental conditions. The following subsections describe the data sources, acquisition methods, and processing techniques in detail.

2.2.1. Flagship Species Monitoring Data

A combined sampling strategy based on fixed plots and line transects was adopted to monitor the breeding and foraging behavior of bar-headed geese and black-necked cranes. A total of 22 systematically distributed sample plots and 32 transects were established across representative habitat types, with approximately 1 km spacing to ensure comprehensive spatial coverage. Field surveys were conducted three times per month from March to September 2023, encompassing the entire breeding and chick-rearing season. The primary objective of this design was to maximize the collection of valid occurrence records reflecting actual species activity, thereby enhancing the spatial accuracy and ecological representativeness of the dataset. Observers traversed each transect on foot using 8–10× binoculars to record species presence, behavioral patterns, and habitat conditions. When key activity areas were encountered outside predefined transects, supplementary off-transect observations were also carried out [32].

2.2.2. Acquisition and Processing of Environmental Variables

With reference to relevant studies [33,34,35], and considering the environmental context of the Longbao Reserve and the habitat preferences of the target species, we selected 11 environmental variables spanning four categories—land cover, vegetation, anthropogenic disturbance, and natural environmental conditions—for use in species distribution modeling (

Table 1. Descriptions and sources of environmental variables used in MaxEnt and Random Forest models for evaluating species habitat suitability.

Category	Description	Code	Spatial Resolution (m)	Year	Source
Land cover	Land-Use Type	fl	0.8	2023	GF-2 (CRESDA, https://database.eohandbook.com/database/agencysummary.aspx?agencyID=130, accessed on 16 May 2024)
	June NDVI	NF	10	2023
Vegetation	NDVI-change1 (11–31 March vegetation dynamics)	N1	10	2023	Google Earth Engine platform (https://earthengine.google.com/, accessed on 16 May 2024)
	NDVI-change2 (31 March–8 April vegetation dynamics)	N2	10	2023
	NDVI-change3 (8–20 April vegetation dynamics)	N3	10	2023
	NDVI-change4 (20 April–13 May vegetation dynamics)	N4	10	2023
	NDVI-change5 (13–25 May vegetation dynamics)	N5	10	2023
	NDVI-change6 (25 May–15 June vegetation dynamics)	N6	10	2023
Natural environmental	Distance to water	Water	0.8	2023	GF-2 (CRESDA, https://database.eohandbook.com/database/agencysummary.aspx?agencyID=130, accessed on 16 May 2024)
Anthropogenic disturbance	Distance to roads	Road	0.8	2023
	Distance to settlements	Settlement	0.8	2023

) [36,37]. Land-use classification data were derived through visual interpretation of GF-2 satellite imagery. Anthropogenic disturbance indicators and water body features were extracted from the classification results. Given the unavailability of spatially explicit livestock density data for the study area, we addressed this gap by incorporating two spatial proxies of grazing pressure—distance to roads and distance to settlements—into the habitat suitability modeling framework. These variables are widely recognized as indirect indicators of pastoral disturbance, especially in alpine wetland–grassland ecosystems, where livestock movements and grazing intensity are closely coupled with road accessibility and village proximity. As such, these proxies allowed us to approximate the spatial pattern of anthropogenic disturbance related to grazing activities in the absence of direct livestock data [15,38,39]. It should be noted that while these proxies represent static spatial attributes, they are appropriate for the Longbao Reserve context, where human activity is minimal and tourism is strictly prohibited. Dynamic anthropogenic pressures, such as seasonal tourism influxes, are negligible within the reserve and thus not included in the modeling framework. Geological variables were not included in the final modeling due to low variability across the study area and limited ecological relevance in this context.

2.3. Methods

This study integrates field-collected avian observations with multi-source environmental datasets from the Longbao Reserve, selecting the black-necked crane and the bar-headed goose as focal species. The assessment of habitat suitability considered key environmental factors, including land use, vegetation conditions, human disturbance, and natural abiotic conditions, within the species’ breeding and foraging areas. Habitat modeling was performed using MaxEnt and Random Forest algorithms, with ensemble integration weighted according to AUC-derived performance metrics. Based on the modeling outputs, a hierarchical weighting scheme was applied to habitat functional types and species distributions to generate integrated habitat suitability maps. Habitat suitability classes were classified using the Jenks Natural Breaks method and overlaid with existing functional zones to refine spatial boundary delineations. This approach produced a scientifically grounded and operationally viable refined zoning scheme, enabling partitioned conservation of high-elevation wetland flagship species and spatially explicit management interventions. The overall research framework is illustrated as illustrated in Figure 2 below.

2.3.1. MaxEnt Model

This study employed the MaxEnt algorithm (MaxEnt software v3.4.4, available at https://biodiversityinformatics.amnh.org/open_source/maxent/, accessed on 13 August 2023) to assess habitat suitability for the breeding and foraging grounds of the black-necked crane and the bar-headed goose. The model was trained using known species occurrence points and 11 environmental variables, with 75% of the data allocated for model training and 25% reserved for testing. Variable importance was evaluated using logistic output and the Jackknife test, while model performance was assessed using the Area Under the Curve (AUC) metric, with AUC values > 0.9 indicating excellent predictive accuracy [40]. To ensure spatial consistency and model compatibility, all environmental variables were standardized in ArcGIS (version 10.8, Esri, Redlands, CA, USA) by clipping to the Longbao Reserve study area, resampling to a 10 m resolution, and converting to ASCII format to meet the input requirements of both MaxEnt and Random Forest models [35,41].

2.3.2. Random Forest Model

This study applied the Random Forest algorithm implemented in R software (v4.3.0, R Foundation for Statistical Computing, Vienna, Austria) to predict habitat suitability for the breeding and foraging grounds of the black-necked crane and the bar-headed goose [38,42,43,44,45]. The dataset was partitioned into 75% training and 25% testing subsets to improve model robustness and generalizability [43]. To ensure robust model generalization, we adopted a standard data-splitting strategy and assessed predictive performance using AUC-based evaluation. This approach enabled consistent comparison across model outputs while minimizing overfitting risks [38,44]. AUC values, ranging from 0 to 1, reflect the area under the Receiver Operating Characteristic (ROC) curve, with higher scores indicating stronger predictive performance. Interpretive thresholds are typically categorized as follows: poor (0.6–0.7), fair (0.7–0.8), good (0.8–0.9), and excellent (0.9–1.0) [44].

2.3.3. Model Ensemble Strategy

To enable comprehensive habitat suitability assessment across multiple species and modeling approaches in high-elevation wetlands, this study implemented a hierarchical stepwise weighting strategy to integrate predictions by species (the black-necked crane and the bar-headed goose), habitat type (breeding vs. foraging grounds), and modeling algorithm (MaxEnt, Random Forest), ultimately producing composite habitat suitability maps for flagship species. Given the pivotal role of breeding habitats in species life cycles and population viability, these areas are generally assigned higher conservation priority than non-breeding sites [39,46,47]. Accordingly, a weighting ratio of 0.6 was assigned to breeding ground predictions and 0.4 to foraging ground predictions, where S_BNC,breed stands for habitat suitability for black-necked crane breeding grounds, S_BNC,habitat stands for habitat suitability for the black-necked crane foraging grounds, S_BHG,breed stands for habitat suitability for the bar-headed goose breeding grounds, and S_BHG,habitat stands for habitat suitability for the bar-headed goose foraging grounds. This weighting scheme was informed by conservation practice [39], as formalized in Equations (1) and (2):

$${\mathrm{S}}_{\mathrm{B}\mathrm{N}\mathrm{C}}=0.6\times {\mathrm{S}}_{\mathrm{B}\mathrm{N}\mathrm{C},\mathrm{breed}}+0.4\times {\mathrm{S}}_{\mathrm{B}\mathrm{N}\mathrm{C},\mathrm{habitat}}$$

(1)

$${\mathrm{S}}_{\mathrm{B}\mathrm{H}\mathrm{G}}=0.6\times {\mathrm{S}}_{\mathrm{B}\mathrm{H}\mathrm{G},\mathrm{breed}}+0.4\times {\mathrm{S}}_{\mathrm{B}\mathrm{H}\mathrm{G},\mathrm{habitat}}.$$

(2)

In Equation (1), S_BNC represents the composite habitat suitability for the black-necked crane, while in Equation (2), S_BHG represents the composite habitat suitability for the black-necked crane.

To generate a unified suitability map across species, we integrated the habitat-weighted outputs of the two target species. Specifically, S_BNC represents the composite suitability for he black-necked crane (as defined in Equation (1)), while S_BHG denotes the corresponding value for the bar-headed goose, calculated analogously. Given that the black-necked crane is classified as a Grade I National Key Protected Animal in China and is globally listed as Vulnerable, it merits higher ecological conservation priority and stronger policy support [25,26]. In contrast, the bar-headed goose has a relatively lower conservation status, but holds numerical dominance among waterbird populations in the Longbao Reserve. Reflecting both species’ protection levels and local ecological roles, weights of 0.7 and 0.3 were assigned to the black-necked crane and the bar-headed goose, respectively. This weighting scheme was informed by species conservation status and follows established methodologies used in multi-species prioritization and ecoregional conservation planning [48,49,50].

$${\mathrm{S}}_{\mathrm{species}}=0.7\times {\mathrm{S}}_{\mathrm{B}\mathrm{N}\mathrm{C}}+0.3\times {\mathrm{S}}_{\mathrm{B}\mathrm{H}\mathrm{G}}.$$

(3)

Equation (3) defines S_species as species-level composite suitability, effectively prioritizing the black-necked crane conservation while incorporating the bar-headed goose requirements.

This study adopted a model ensemble weighting strategy that integrated the predictive performance of both MaxEnt and Random Forest across species and habitat scenarios. Separate modeling was conducted in both MaxEnt and Random Forest for the two habitat types (breeding and foraging) of each species (the black-necked crane and the bar-headed goose), with corresponding AUC metrics obtained for each model. The AUC values across the four habitat scenarios were summed separately to obtain the total AUC for MaxEnt (AUC_ME) and Random Forest (AUC_RF). For model integration, weights were calculated using Equation (4) based on AUC performance metrics [51,52]:

$${\mathsf{\omega }}_{\mathrm{M}\mathrm{E}}={\displaystyle \frac{\mathrm{A}\mathrm{U}{\mathrm{C}}_{\mathrm{M}\mathrm{E}}+\mathrm{A}\mathrm{U}{\mathrm{C}}_{\mathrm{R}\mathrm{F}}}{\mathrm{A}\mathrm{U}{\mathrm{C}}_{\mathrm{M}\mathrm{E}}}}, {\mathsf{\omega }}_{\mathrm{R}\mathrm{F}}={\displaystyle \frac{\mathrm{A}\mathrm{U}{\mathrm{C}}_{\mathrm{M}\mathrm{E}}+\mathrm{A}\mathrm{U}{\mathrm{C}}_{\mathrm{R}\mathrm{F}}}{\mathrm{A}\mathrm{U}{\mathrm{C}}_{\mathrm{R}\mathrm{F}}}}$$

(4)

Notably, the Random Forest model exhibited reduced predictive accuracy for the black-necked crane (AUC: 0.70–0.78; Section 3.1), likely due to its sparse population density (175 individuals) and limited occurrence records relative to the reserve’s spatial extent. This limitation aligns with known challenges of tree-based algorithms in modeling rare species with small sample sizes [52,53]. To mitigate potential bias from single-model dependencies, our AUC-weighted ensemble explicitly prioritized higher-performing models while retaining complementary insights from Random Forest [51]. Here, S_ME and S_RF represent the species-level suitability outputs from the MaxEnt and Random Forest models, respectively, and the final composite suitability is computed via Equation (5):

$${\mathrm{S}}_{\mathrm{final}}={\mathsf{\omega }}_{\mathrm{M}\mathrm{E}}\times {\mathrm{S}}_{\mathrm{M}\mathrm{E}}+{\mathsf{\omega }}_{\mathrm{R}\mathrm{F}}\times {\mathrm{S}}_{\mathrm{R}\mathrm{F}}.$$

(5)

In Equation (5), S_final denotes the final composite habitat suitability prediction for flagship species generated in this study.

2.3.4. Delineation of Refined Zoning

Habitat suitability outputs were classified into four categories—highly suitable, suitable, moderately suitable, and unsuitable—using the Jenks Natural Breaks Classification method in ArcGIS [24,54]. Guided by the principles of suitability prioritization, functional zone compatibility, and spatial refinement, a comprehensive reclassification was conducted by integrating habitat suitability levels with the spatial overlay of existing management zones. In core zones, highly suitable and suitable areas were designated as Core Habitat Conservation Zones, moderately suitable areas as Ecological Rehabilitation Zones, and unsuitable areas as Ecological Management Zones [55]. In experimental zones, moderately suitable areas were designated as Ecological Management Zones, while unsuitable areas were classified as Sustainable Utilization Zones. These zones primarily retain existing land-use patterns and are managed with a focus on minimizing human disturbance [56,57]. In buffer zones, highly suitable and suitable areas were similarly designated as Core Habitat Conservation Zones, moderately suitable areas as Ecological Management Zones, and unsuitable areas as Sustainable Utilization Zones. The final zoning framework incorporated considerations of patch size and management feasibility, merging fragmented patches, and smoothing zone boundaries to generate a scientifically robust and operationally feasible fine-scale zoning pattern [58].

3. Results

This section presents the outcomes of our integrated modeling framework, which assessed habitat suitability for flagship species to inform adaptive zoning in the Longbao Reserve. We first evaluate model performance across algorithms and species, then characterize the spatial distribution of habitat suitability, identify key environmental drivers, and finally propose a refined zoning scheme based on ecological congruence.

3.1. Model Evaluation Results

The performance evaluation of the models demonstrated strong predictive capabilities, particularly for the bar-headed goose. As shown in

Table 2. Model prediction accuracy.

Species and Habitat Type	MaxEnt (AUC)	RF (AUC)
Bar-headed goose breeding site	0.99	0.99
Bar-headed goose foraging site	0.94	0.91
Black-necked crane breeding site	0.91	0.78
Black-necked crane foraging site	0.96	0.70

, both MaxEnt and Random Forest achieved high AUC values for this species’ breeding habitats (MaxEnt: 0.99; RF: 0.99) and foraging habitats (MaxEnt: 0.94; RF: 0.91), indicating stable habitat preferences and the availability of high-quality training data. In contrast, for the black-necked crane, although MaxEnt performed well in both breeding (0.91) and foraging (0.96) predictions, Random Forest exhibited lower predictive accuracy (0.78 and 0.70, respectively), likely due to limited sample sizes and greater ecological heterogeneity. These AUC results informed the subsequent ensemble weighting strategy, ensuring that higher-performing models contributed proportionally more to the final habitat suitability outputs. The consistently strong performance of MaxEnt with small sample sizes further reinforces its appropriateness for species such as the black-necked crane, whose distribution is characterized by spatial sparsity and fragmentation.

3.2. Spatiotemporal Distribution of Habitat Suitability

With respect to the spatial distribution of habitat suitability for flagship species, both the MaxEnt and Random Forest models demonstrated strong overall agreement in spatial prediction outcomes (Figure 3), although each model exhibited distinct strengths in capturing different levels of habitat suitability. For the bar-headed goose, the MaxEnt model identified highly suitable habitats with a markedly clustered distribution and sharply defined boundaries, highlighting the spatial concentration and clarity of core habitat zones. In contrast, the Random Forest model depicted a broader extent of moderately suitable areas, revealing greater habitat connectivity and offering improved discrimination in transitional zones. For the black-necked crane, the MaxEnt model produced relatively intact clusters of a highly suitable habitat, reinforcing core habitat characteristics. Conversely, the Random Forest model yielded a more fragmented distribution in moderate and low suitability zones, with smoother spatial transitions that more accurately reflected the heterogeneous nature of underlying habitat conditions. The integrated results for flagship species showed strong consistency in overall spatial patterns across both models, with highly suitable core habitats concentrated in the wetland-dominated central and eastern regions of the study area.

Figure 4 illustrates the spatial heterogeneity in the distribution of integrated habitat suitability classes. Highly suitable zones were concentrated around lakes and adjacent wetlands in the central to southeastern parts of the reserve, comprising 7.9% of the total area. Moderately suitable zones (9.1%) were primarily located along wetland margins and within marshland sectors. Low suitability zones (16.9%) were concentrated in lakeshore extensions and gently sloping areas. Unsuitable zones accounted for the largest proportion (66.1%) and were predominantly distributed in the northern and southwestern sectors of the reserve. Overall, highly and moderately suitable zones formed spatially contiguous clusters, whereas low suitability and unsuitable zones exhibited broader spatial coverage with annular or peripheral distribution patterns. Notably, the spatial overlay of habitat suitability with functional zones (as elaborated in Section 3.4) revealed significant mismatches—with several highly suitable habitats located outside designated strict protection zones, highlighting the need for refined zoning strategies.

3.3. Key Environmental Factors and Their Response Curves

To systematically evaluate the relative importance of environmental variables in predicting species habitat suitability, this study employed both MaxEnt and Random Forest models to generate predictions and derive variable importance metrics. Specifically, MaxEnt utilized permutation importance, while Random Forest employed a Gini index-based measure (Mean Decrease Gini) [41]. To account for performance differences across modeling locations, variable importance values were integrated using a weighting scheme based on average test AUC scores across four modeling sites (MaxEnt = 0.895; RF = 0.850) [38]. This weighting strategy reflected model-specific explanatory power and reduced potential bias from reliance on a single model, thereby enhancing the robustness of the results. Based on the integrated analysis, the top four most influential environmental variables were identified for each species and habitat stage, representing the dominant environmental drivers [59]. Variable importance was extracted from both MaxEnt and Random Forest models and weighted by model AUC scores to evaluate environmental influence on habitat suitability predictions (Figure 5). Distance to roads, settlements, and water, along with land-use type, emerged as the key contributing variables. Among them, land-use type (fl) was the most influential, contributing over 30% to the habitat suitability models for both the breeding and foraging habitats of the bar-headed goose, as well as the foraging habitat of the black-necked crane. The NDVI in June (nf) was particularly important for the breeding habitat of the black-necked crane, contributing more than 35%. Additionally, distance to water, roads, and settlements demonstrated consistent influence across habitat stages. In contrast, NDVI variation metrics (n1–n6) exhibited relatively low overall contributions, suggesting limited impact on habitat suitability predictions. Although June NDVI emerged as a dominant variable, our findings also highlight the substantial contributions of other predictors such as land-use type and proximity to water bodies, roads, and settlements. This suggests that habitat suitability is shaped by a diverse array of environmental drivers beyond vegetation greenness alone.

Figure 6 shows that the foraging habitat of the black-necked crane is strongly associated with lake surfaces and wetland ecosystems, and exhibits a positive relationship with June NDVI, as suitability increases markedly with higher vegetation greenness. Suitability declines sharply with increasing distance from water bodies, indicating a strong dependence on water availability during foraging. Foraging sites also show a clear preference for areas located farther from roads, suggesting selection for environments with minimal human disturbance. Breeding habitats of the black-necked crane similarly exhibit a strong preference for lake and wetland ecosystems, with habitat suitability positively correlated with June NDVI. The species demonstrates marked reliance on vegetation-stable zones, particularly those characterized by NDVI-change4 (20 April–13 May) and NDVI-change5 (13 May–25 May). Suitability decreases significantly under conditions of either rapid greening or vegetation loss, indicating a preference for habitats with relatively stable vegetation dynamics. Breeding sites are predominantly located near water bodies and farther from roads, highlighting dual preferences for proximity to water and low anthropogenic disturbance. Breeding habitat suitability for the bar-headed goose is primarily shaped by land-use and vegetation dynamics during mid-to-late spring, with a distinct preference for humid habitats such as wetlands and lakes. This is especially evident in their selection of rapidly greening areas during 20 April–13 May (NDVI-change4). Suitability peaks at approximately 1500 m from roads, suggesting a tendency to avoid areas of high human activity. A clear positive trend is also observed with an increasing June NDVI. Foraging habitats are frequently situated at the interface between wetlands and bare land, with a preference for high NDVI regions, indicating a tendency to forage in areas with lush vegetation. Suitability declines with increasing distance from water bodies, confirming strong water dependence. Additionally, suitability follows a “V”-shaped relationship with distance to settlements, peaking at intermediate distances (~2000 m), suggesting avoidance of densely populated areas while potentially benefiting from edge-associated ecological resources.

3.4. Refined Management Zoning

Building upon the existing functional zoning framework of Longbao Reserve, management units were delineated by integrating flagship species’ habitat suitability with key environmental drivers, followed by vector-based mapping of all functional sub-zones to support science-based management of the protected area. The corresponding zoning results are illustrated in Figure 7.

(1)

Core Habitat Conservation Zones accounted for 16.9% of the study area. These zones were primarily located in undisturbed lakes and wetlands, exhibiting peak ecological suitability, with minor extensions into the margins of buffer zones. Collectively, these areas meet the habitat requirements of flagship species through optimal land cover, proximity to water, healthy vegetation conditions, and minimal anthropogenic disturbance.

(2)

Ecological Rehabilitation Zones comprised 7.2% of the study area. These were situated along the periphery of core zones and extended into buffer areas. Characterized by lower habitat suitability scores, they still fulfill the basic habitat needs of bar-headed geese and black-necked cranes. Well-developed vegetation and moderate levels of human pressure make these areas suitable for targeted ecological restoration.

(3)

Ecological Management Zones covered the largest proportion (53.5%) of the study area, mainly distributed across western experimental zones and southern buffer zones. These zones exhibited low-to-moderate habitat suitability, intermediate NDVI values, average vegetation coverage, and moderate habitat quality. Given their proximity to roads and settlements, they are appropriate for maintaining ecological continuity through disturbance mitigation measures.

(4)

Sustainable Utilization Zones accounted for 22.4% of the study area and were distributed across central experimental zones and peripheral buffer areas in the north and west. These areas are characterized by low vegetation cover, considerable distance from water bodies, and high levels of anthropogenic disturbance. Although they exhibit the lowest ecological suitability, they remain appropriate for controlled use and development under strict ecological safeguards.

Spatial overlay analysis between habitat suitability classifications and the original secondary-level functional zones (Figure 7) indicates that, although the original core zone was designated as the highest-level protection area overall, significant ecological heterogeneity exists within its boundaries. Certain sections were reclassified as Ecological Management Zones, suggesting that even within priority conservation areas, some sites exhibit lower ecological suitability or elevated levels of anthropogenic disturbance, necessitating context-specific and differentiated management approaches. These findings underscore the importance of adaptive management strategies to accommodate ecological variability within protected areas. In buffer zones, management tiers were progressively arranged from the periphery toward the core, forming clear ecological transition gradients that enhance buffering capacity. In experimental zones, however, some wetland sectors were reassigned as Ecological Management Zones or even Core Habitat Conservation Zones, revealing that ecologically valuable habitats were previously overlooked in the original zoning scheme. Collectively, this refined adjustment corrected functional mismatches and improved alignment with actual ecological functions, while preserving the core structure of the original zoning framework—thereby facilitating a shift from extensive to precision-based ecological management.

4. Discussion

4.1. Suitability Distribution and Influencing Factors

Notwithstanding sample size constraints for the endangered black-necked crane, our AUC-weighted ensemble framework effectively leveraged the complementary strengths of MaxEnt (suitable for sparse samples) and Random Forest (effective for modeling complex interactions). This approach ensured robust habitat suitability predictions while transparently addressing the data limitations inherent in rare species. The research model evaluation revealed notable performance discrepancies between the two algorithms in predicting habitat suitability for flagship species. The comparatively lower predictive accuracy of the Random Forest model for black-necked cranes (AUC: 0.70–0.78) likely stems from their low population density (175 individuals in 2023) and restricted distribution range within the reserve. These factors constrained the model’s ability to capture fine-scale habitat heterogeneity, particularly given Random Forest’s reliance on larger sample sizes for effective feature selection. While the AUC-weighted ensemble approach mitigated this limitation by emphasizing the more reliable MaxEnt outputs (AUC: 0.91–0.96), future studies would benefit from expanded field monitoring to improve sample representativeness for rare species [38,44,46]. In contrast, the MaxEnt model demonstrated greater robustness under small-sample conditions, delivering higher predictive accuracy and reliability for both the breeding and foraging habitats of black-necked cranes [35]. To further address model discrepancies, an AUC-weighted ensemble strategy was adopted, assigning greater weight to higher-performing models to enhance the scientific validity and robustness of the integrated predictions [51,52,53].

Modeling results indicate that both bar-headed geese and black-necked cranes exhibit pronounced preferences for aquatic and wetland habitats. This spatial pattern aligns with the environmental determinism framework described by Wang et al. (2022) and Fu et al. (2023) in their studies of waterbird distributions across the Qinghai–Tibet Plateau’s wetland ecosystems [34,60]. Unsuitable habitats are primarily located in grassland areas adjacent to wetlands and in regions with intensive human activity, influenced by three primary limiting factors: water scarcity, low vegetation quality, and high levels of anthropogenic disturbance. This pattern is consistent with the findings of Mi et al. (2017) regarding spatial habitat modeling for Asian crane species [38]. Vegetation conditions and seasonal dynamics emerged as key ecological drivers shaping the habitat suitability of these flagship species. Bar-headed geese are highly responsive to June NDVI values and springtime vegetation dynamics, exhibiting a distinct preference for areas showing moderate increases in vegetation greenness from 20 April to 13 May, during the early breeding season. This reflects ecological adaptations to aquatic vegetation and early-stage grassland development [33,35]. Black-necked cranes demonstrate higher breeding suitability in zones characterized by stable vegetation conditions, particularly during the critical reproductive period from 20 April to 25 May, indicating a preference for ecologically stable breeding habitats. This finding aligns with the theory proposed by Suttidate et al. (2023) regarding “the determining role of microhabitat heterogeneity in breeding site selection” [53]. During the non-breeding period, black-necked cranes respond positively to absolute NDVI levels, with peak suitability occurring at approximately NDVI = 0.4, suggesting a preference for areas with relatively high vegetation biomass during this stage.

Overall, habitat suitability patterns in Longbao Wetland are driven not only by static factors such as land use and proximity to water bodies, but also by seasonal vegetation dynamics. The incorporation of multi-temporal NDVI metrics significantly enhanced the models’ sensitivity to environmental seasonality, corroborating insights from related research [24]. These findings provide crucial ecological evidence for understanding avian habitat selection and spatial distribution in alpine wetland ecosystems.

4.2. Precision Zoning and Management Control

Established as China’s first high-elevation wetland reserve designated for the breeding protection of black-necked cranes, Longbao Nature Reserve has long struggled with coarse-scale zoning and poorly defined management boundaries. Its current functional zoning has proven inadequate in alleviating human–wildlife conflicts within the reserve. Given Longbao’s role as a representative high-altitude wetland providing a critical habitat for rare waterbirds, including black-necked cranes and bar-headed geese, the existing zoning framework requires refinement through functional realignment and differentiated management strategies in peripheral zones. This integrated approach safeguards essential waterbird habitats while allowing for balanced spatial allocation to support pastoral livelihoods, thereby alleviating tensions between community development and ecological conservation.

Wetland and grassland ecosystems in Longbao exhibit distinct spatial distributions across the landscape, with clear differentiation between natural areas and zones influenced by human activity. This spatial contrast provides a natural foundation for implementing differentiated management strategies. Based on the spatial distribution of habitat suitability for flagship species and gradients of human disturbance, this study developed a multi-level functional zoning scheme accompanied by tailored management strategies to enable spatially targeted governance (Figure 8). The Core Habitat Conservation Zone encompasses key breeding and foraging areas for flagship species, where all grazing activities are strictly prohibited. In the Ecological Rehabilitation Zone, rotational grazing should be implemented, with grazing prohibited during critical bird breeding periods to minimize disturbance to nesting and chick-rearing. This approach aims to improve ecological quality and expand suitable habitat areas. During the non-breeding season, moderate rotational grazing may be permitted based on ecological monitoring, to manage vegetation structure, suppress the overgrowth of dominant species, and support ecosystem stability and community diversity [61]. In the Ecological Management Zone, frequent grazing activities have significantly impacted the foraging behavior of flagship species. Persistent disturbances should be mitigated by limiting grazing intensity and applying rotational systems with defined seasonal timeframes and cycles, thereby reducing interference during species activity periods [62]. The Sustainable Utilization Zone, designated primarily for human use, should restrict high-intensity continuous grazing and limit grazing along wetland margins, while managing livestock movement in alignment with community development goals. In response to rising ecotourism demand and existing infrastructure, management should promote low-impact human activities. Herders should be encouraged to participate in ecological interpretation, environmental education, and eco-product demonstration, thereby diversifying livelihood options, enhancing ecological stewardship, and supporting the integrated development of conservation and pastoral economies [63]. Future implementation should consider institutionalizing incentive mechanisms, such as ecological compensation, payments for ecosystem services (PESs), and conservation stewardship agreements, to enhance community engagement and improve the perceived legitimacy of zoning policies. Specifically, ecological compensation schemes may target herder households the grazing access of which is restricted due to zoning regulations. Compensation could take the form of direct financial subsidies, rotational pasture access, or employment opportunities in conservation programs, such as ecological ranger positions. PES schemes can also be designed around the wetland’s ecosystem services—such as water retention or carbon sequestration—with funding sourced from downstream beneficiaries, government agencies, or third-party conservation organizations. To ensure equity and local buy-in, these mechanisms should be co-designed with community input and supported by transparent governance structures. Such incentive-based approaches have been shown to improve compliance and promote the long-term sustainability of protected area management [63,64,65].

Currently, many protected areas continue to face challenges stemming from overly generalized zoning frameworks and inflexible management practices, which often lack sufficient consideration of species-specific habitat requirements and patterns of human activity, thereby exacerbating human–wildlife conflicts [64]. In contrast, this study integrates species distribution patterns with gradients of anthropogenic disturbance to establish adjustable, multi-tiered management units, enabling differentiated and responsive management strategies. This approach emphasizes the adaptive adjustment of management intensity based on localized ecological and social conditions, striking a balance between ecological conservation objectives and community needs, and demonstrates strong potential for broader application. The refined and differentiated zoning framework proposed in this study can be adaptively implemented across other national parks and protected areas and tailored to specific local contexts, thereby enhancing the scientific rigor and adaptive management capacity of protected area governance.

4.3. Limitations and Prospects

Caution is needed in generalizing this framework to all high-altitude ecosystems. This study focused only on two waterbird species in a marsh–grassland setting. Broader applicability should be tested using other taxa such as mammals and amphibians, and other habitats like montane forests and alpine tundra. Though integrating habitat modeling and zoning for flagship species, the analysis lacks long-term data such as population structure, reproductive success, and density. Annual changes in abundance, behavior, or migration require periodic reassessment. Multi-season monitoring and dynamic models are recommended. Key environmental indicators like soil moisture and water quality—such as chlorophyll-a, turbidity, and dissolved oxygen—were not included. These factors, especially during breeding and foraging, could enhance model accuracy beyond the NDVI.

The study area—a small, geomorphologically uniform alpine wetland—limits generalizability. Although sampling was systematic, occurrence records were few relative to the reserve’s size. This study thus serves as a field-constrained methodological trial. Application to broader, heterogeneous wetlands with complex human–wildlife dynamics, such as those across the Qinghai–Tibet Plateau, should be explored in future work. Human disturbance was considered but lacked spatial quantification of grazing. Incorporating high-resolution land-use and remote sensing indicators would improve accuracy. Ensemble modeling reduced algorithm bias, but Random Forest yielded lower AUC values for black-necked cranes, likely due to sparse data. Long-term monitoring, telemetry data, and spatially explicit models like integrated nested Laplace approximations are needed for rare species.

5. Conclusions

This study developed ensemble species distribution models by integrating MaxEnt and Random Forest algorithms, using field-observed breeding and foraging data and key environmental variables for the bar-headed goose and the black-necked crane. The models revealed spatial patterns of habitat suitability, supporting the design of a refined functional zoning scheme that meets both ecological conservation and community development needs. The modeling framework achieved strong predictive performance (AUC: 0.84–0.93), with land-use type, June NDVI, and proximity to water and roads identified as the main environmental drivers. Approximately 47% of the reserve was classified as moderately to highly suitable for the target species. Based on classified habitat suitability and ecological considerations, a four-level zoning plan was established, including Core Habitat Conservation Zones, Ecological Rehabilitation Zones, Ecological Management Zones, and Sustainable Utilization Zones. This structure provides a graded management framework, enabling the prioritization of high-value habitats while allocating designated areas for grazing and reducing spatial conflicts through clearly defined land-use boundaries.

The proposed approach is particularly applicable to protected areas characterized by clear flagship species–habitat associations, such as plateau wetlands, where species activities are spatially concentrated and habitat boundaries are distinct. It provides a replicable pathway for implementing adaptive, species-centered conservation planning that balances ecological priorities with human land-use demands.