Conservation Effectiveness and Heterogeneity of the National Park in Promoting Ecosystem Health: Causal Evidence from Huangshan, China

Abstract

National parks are key tools for safeguarding ecosystem health, yet their conservation performance remains unclear. Comprehensive evaluations are crucial for guiding targeted and effective conservation strategies. This study employed the Vigor–Service–Resilience (VSR) framework together with causal inference models to assess the role of Huangshan National Park (HNP) in promoting ecosystem health and to examine the heterogeneity of its ecological outcomes from 2010 to 2020. The results indicate that (1) ecosystem health improved significantly across the region, with 69.5% of pixels showing positive change, particularly in ecosystem services and vigor; (2) compared with matched unprotected sites, HNP enhanced EH by 5.7% in 2010, 3.4% in 2015, and 6.5% in 2020, and also generating positive spillover effects within 30 km of its boundaries; (3) conservation impacts differed notably across socio-ecological conditions, with greater benefits observed in areas of lower elevation, gentle slopes, and reduced precipitation. These findings provide robust causal evidence of the protective value of HNP and underscore the importance of targeted and cost-efficient management strategies to optimize conservation outcomes and support sustainable regional development.

1. Introduction

Under the 2030 Agenda, protected areas are regarded as essential for achieving multiple Sustainable Development Goals (SDGs), playing a positive role in sustaining biodiversity [1] and enhancing ecosystem services [2]. Over the past 40 years, China has experienced a rapid expansion of its protected area network, now covering over 17% of its land and 3.5% of its marine territory—meeting the Aichi Target of 17% by 2020 [3,4]. To better protect the core natural landscape and respond to the “eco-civilization” initiative, China launched a series of pilot national parks [5], aiming to protect ecosystem integrity while also supporting community development and human well-being [6]. Yet, simply enlarging these sites does not guarantee effective ecosystem protection [3]. Many protected areas have been labeled as “paper parks” due to inadequate management resources, while others face mounting threats from local development and urbanization pressures [7,8]. These challenges highlight the need for systematic and scientific evaluations to determine whether protected areas are truly achieving their intended conservation goals.

The concept of ecosystem health has been introduced as a central solution to address ecological problems and is now widely recognized and applied around the world [9,10]. A healthy ecosystem reflects its ability to remain stable and provide essential services at specific spatial and temporal scales [10], which is vital for ensuring access to natural resources and for supporting public health, economic stability, and social well-being [11]. Many studies have adopted biological indicators or comprehensive indicator frameworks to assess ecosystem health [12], which are important for ecosystem conservation and monitoring. Previous bio-centric evaluations used species indicators such as microcystins [13], butterflies [14], or birds [15] to represent ecosystem health, while overlooking the impact of widespread human activities, especially in China’s protected areas. Severe human–land conflicts make the traditional “fortress conservation” approach both ecologically inappropriate and practically unfeasible [16]. On the other hand, many studies have developed multi-dimensional frameworks to evaluate ecosystem health. Examples include the hierarchical analysis method [17], the Pressure–State–Response (PSR) framework [18], the Drivers–Pressures–State–Impact–Response–Management (DPSIRM) framework [19], and the Vigor–Organization–Resilience (VOR) framework [20]. However, these approaches often overlook ecosystem services linked to human well-being, as national parks are expected not only to maintain the integrity of natural ecosystems but also to improve local livelihoods and provide opportunities for research, education, and tourism [21]. To address this gap, scholars proposed the vigor–services–resilience (VSR) framework based on the unique and sensitive human–nature relationship in national parks [16,22]. It is gaining recognition as an effective methodology, emphasizing the importance of biodiversity conservation, the role of ecosystem services in improving community well-being, and the resilience needed to maintain stability and ensure sustainability under disturbance [23].

Assessing conservation effectiveness is essential to determine whether protected areas achieve their intended goals. Previous studies have examined this topic from perspectives such as funding inputs or reductions in human and environmental threats [24,25], but these approaches fail to show a clear causal link between ecosystem health and management actions. Since many protected areas are in remote regions with little human disturbance and low social costs, ecosystems there may stay healthy even without formal protection, which can make evaluations misleading [7,26]. To reduce the bias from non-random site selection [27], researchers increasingly recognize the importance of using causal inference models to identify matched unprotected areas as counterfactuals [28]. Moreover, although growing attention has been given to the outcomes of conservation policies, few studies have connected this evidence to spatially explicit protection priorities that can inform region-based planning [29]. Even fewer have explored how spatial differences in conservation outcomes can be used to create suitability maps for guiding such priorities [30]. Examining the heterogeneity of national park effectiveness and applying this knowledge is crucial for setting more efficient future conservation strategies. Such evaluations are also vital for optimizing the quality of China’s national park system and for contributing to global conservation goals such as the United Nations 30×30 target.

To address the identified limitations, this study conducts an empirical analysis of Huangshan National Park (HNP) to examine its conservation effectiveness on ecosystem health. As a World Biosphere Reserve and a UNESCO World Cultural and Natural Heritage Site, the park is a model for sustainable management [31]. By combining rigorous counterfactual analysis with remote sensing data, this study aims to: (1) measure the spatial distribution and temporal changes in ecosystem health in HNP from 2010 to 2020; (2) evaluate the effects of HNP on enhancing ecosystem health using propensity score matching to build comparison groups from nearby unprotected areas; and (3) identify the heterogeneity in conservation outcomes across different social and environmental settings to determine priority areas for protection.

2. Materials and Methods

2.1. Study Area

The study area is located in northern Huangshan City, Anhui Province, southeastern China (Figure 1). Its spatial range includes Huangshan Scenic Area, Jiulongfeng Nature Reserve, Wuxishan Nature Reserve, Tianhu Nature Reserve, and the ecological corridors connecting these protected areas, covering a total area of approximately 330.3 km² [32]. As one of the candidate areas for China’s priority national park construction [33,34], Huangshan is world-famous for its granite geology and spectacular landforms [32]. It also serves as the core zone of the World Biosphere Reserve [31], and is home to 1805 native plant species and 323 wildlife species, including 28 under national protection such as the Chinese merganser and the clouded leopard [35,36].

Historically, Huangshan was relatively underdeveloped until modern tourism began in 1979 [31,37]. Since then, tourism has grown rapidly, and most residents have become involved in the industry [38]. By 2024, the park received 5.57 million visitors and generated 3.45 billion RMB in revenue [39]. Green practices such as “clean supplies go up, waste comes down” and the “closed rotation” management system have helped restore forests while keeping balance with use [40]. These efforts earned Huangshan a place on the IUCN Green List for its achievements in ecosystem management and cultural landscape protection [41].

The existence of protected areas can have a spatial spillover effect or a leakage effect on the unprotected areas around them [7,42], which may compromise the accuracy of co-conservation effectiveness evaluations [43]. Particularly, as critical transition zones between national parks and surrounding urban areas, gateway communities exhibit a symbiotic relationship with parks through transportation hubs, sustainable agriculture, environmental education, and eco-recreation [44,45]. As a result, we generated a series of 10 km concentric buffers within a 50 km zone surrounding the park boundary (Figure 2). This approach enables a clear determination of whether conservation effects exist in HNP, and to what extent spillover or leakage effects occur.

2.2. Data Acquisition and Preprocessing

Table 1. Data sources.

Data Name	Data Source	Resolution	Data Usage
The average market value of major crops	National Farm Product Cost–benefit Survey from the National Development and Reform Commission. https://www.ndrc.gov.cn/fgsj/ (accessed on 5 March 2025)	Yearbook data	Qualification of ecosystem health
Net Primary Productivity (NPP)	Resource and Environmental Science and Data Platform of the Chinese Academy of Sciences http://www.resdc.cn/ (accessed on 21 October 2024)	30 m	Qualification of ecosystem health
Land use data	Resource and Environmental Science and Data Platform of the Chinese Academy of Sciences http://www.resdc.cn/ (accessed on 21 October 2024)	30 m	Qualification of ecosystem health
Elevation	Geospatial data cloud https://www.gscloud.cn/ (accessed on 21 October 2024)	30 m	Counterfactual analysis of conservation effects and heterogeneity
Slope	Geospatial data cloud https://www.gscloud.cn/ (accessed on 21 October 2024)	30 m	Counterfactual analysis of conservation effects and heterogeneity
Road data	OpenstreetMap https://www.openstreetmap.org (accessed on 1 February 2025)	Vector data	Counterfactual analysis of conservation effects and heterogeneity
Mean annual precipitation	China Meteorological Data Network http://data.cma.cn/ (accessed on 21 October 2024)	1000 m	Counterfactual analysis of conservation effects and heterogeneity

provides the datasets used in this study. This study incorporated several types of data to quantify the ecosystem health of HNP and its surrounding zones, and applied causal inference models for counterfactual analysis. Specifically, land use, vegetation, and yearbook data were used to calculate ecosystem health, while meteorological, topographic, and transportation data supported the counterfactual evaluation. To ensure consistency across sources, all raster data were reprojected to the WGS 1984 coordinate system. A 30 × 30 m fishnet was generated using the Fishnet tool in ArcGIS 10.8 for data resampling and spatial joining.

2.3. Methods

2.3.1. Ecosystem Health Assessment

Ecosystem health was evaluated through three main dimensions: Ecosystem Vigor (EV), Ecosystem Services (ES), and Ecosystem Resilience (ER). Traditional methods of ecosystem health assessment often overlook either the stability and integrity of internal structures or the impacts of human disturbance [10,46]. Considering that national parks should take into account both natural ecosystems protection while also providing essential services [21], this study integrated both biophysical and human factors [23], and adopted the Vigor–Services–Resilience (VSR) framework for assessing ecosystem health and tracking its variations. According to [10], the model is expressed as:

$$EH=\sqrt[3]{EV\times ES\times ER}$$

(1)

$EH$ represents ecosystem health, while $EV$, $ES$, and $ER$ correspond to ecosystem vigor, services, and resilience. Because the evaluation indicators differ in both units and magnitudes, data from different years are not directly comparable. To resolve this issue, the study applied a global normalization approach, standardizing the indicators across multiple time periods and adjusting all components to a uniform scale of [0, 1].

(1)

Ecosystem Vigor (EV)

Ecosystem vigor refers to the productivity of an ecosystem and shows how active and resilient it is [10]. It captures vegetation performance under nutrient supply, water availability, and broader environmental settings. Greater vigor suggests stronger growth and higher stability. In this study, NPP was adopted as the measure of EV because it reliably reflects photosynthetic efficiency and the ability to store carbon [47,48].

(2)

Ecosystem Services (ES)

Ecosystem services were assessed using the value transfer method developed by Xie et al. (2017) [49]. This approach is widely applied in studies to estimate the spatial and temporal distribution of ecosystem service values. Under this framework, ecosystems are grouped into six main categories: farmland, forest, grassland, wetland, barren land, and water bodies. Their services are further classified into four types: provisioning, regulating, habitat, and cultural or amenity services. Each ecosystem service was assigned an economic value by multiplying an equivalence coefficient with a standard equivalence factor (RMB/ha) [49]. The coefficient is dimensionless, while the factor provides a monetary reference. The factor is defined as the value of natural grain yield per hectare of farmland, based on the national average production. The standard equivalence factor for food production services can be calculated according to [49] as follows:

$$E_a={\displaystyle \frac{1}{7}}\sum _{i=1}^n\left(P_i\times Q_i\right)$$

(2)

$E_a$ denotes the equivalent factor for food production services per unit area (RMB/ha), $i$ refers to the main crops in the study area, $P_i$ indicates the average market price of these crops (RMB/kg), while $Q_i$ represents their yield per unit area (kg/ha). The value of ecosystem services was then estimated with the following formula [50]:

$$S=\sum _{t=1}^n\left(A_t\times {VC}_t\right)$$

(3)

$S$ represents the total ecosystem service value across all ecosystem types. For each ecosystem type $t$, $A_t$ is the land area (ha), and ${VC}_t$ is the value coefficient (RMB/ha·year) for the ecosystem type $t$, which indicates the average service value per hectare per year.

(3)

Ecosystem Resilience (ER)

Ecosystem resilience refers to the capacity of an ecosystem to endure external disturbances. It also reflects its ability to reduce damage through self-regulation and to keep both structure and function stable [10]. The idea of resilience covers two aspects: resistance to stress and the ability to recover afterward. This study applied the equation according to previous research for resilience evaluation [10,51]:

$$R=\left(w_1\times \sum _{i=1}^n{T}_i\times C_{ti}\right)+\left(w_2\times \sum _{i=1}^n{T}_i\times C_{li}\right)$$

(4)

$R$ represents ecosystem resilience, $T_i$ is the area of land type $i$, and $n$ is the total number of land types, $C_{ti}$ indicates the resistance coefficient of land type $i$, $C_{li}$ stands for its resilience coefficient. The weights of resistance and resilience depend on whether external pressures exceed the ecosystem’s self-regulation capacity [52]. Resistance should be emphasized when external disturbances are weaker than the ecosystem’s self-adjustment ability, while resilience becomes more important when pressures are stronger [53]. Unlike urban areas that experience more intensive human activities, Huangshan National Park (HNP) has largely preserved its natural state. Therefore, we assigned a higher weight of 0.6 ($w_1$) to the resistance coefficient ($C_{ti}$), while the resilience coefficient ($C_{li}$) was given a lower weight of 0.4 ($w_2$) according to previous studies [54,55].

In addition, diverse land use types show different resistance and resilience coefficients. In this study, the values for Huangshan National Park and the surrounding areas were taken from the literature [51] (

Table 2. The ecosystem resilience coefficient of each landscape type.

Landscape Type	Grassland	Cropland	Forestland	Construction Land	Water	Unused Land
$C_{li}$	0.8	0.3	0.6	0.2	0.7	0.1
$C_{ti}$	0.6	0.5	1.0	0.3	0.8	0.2

).

2.3.2. Propensity Score Matching for Conservation Effectiveness Assessment

Propensity score matching (PSM) is a widely accepted approach for evaluating the conservation outcomes of protected areas [56]. This approach establishes a robust counterfactual framework to address concerns that observed improvements in ecosystem health may simply result from the protected areas’ location in less disturbed landscapes, rather than from the actual effects of conservation policies. In this study, we applied PSM to identify comparable unprotected pixels that exhibit similar environmental features to protected pixels within the national park (treatment group). Drawing on previous studies [10,57], we controlled for key confounding variables, including altitude, slope, precipitation, population density, and distance to roads. We created a 30 × 30 m fishnet to extract pixel values and confined control pixels to a 50 km buffer around the park boundaries to ensure spatial comparability.

In selecting the matching method, we first tested several matching approaches, including one-to-one nearest neighbor, optimal, full, and subclassification matching. Balance was then tested using standardized mean differences (SMDs) of individual covariates. After adjustment, SMDs for most variables were below 0.25 (

Table A1. Standardized mean differences (SMDs) of covariates before and after alternative matching approaches.

a. 2010
	1:1 Nearest Matching		1:2 Nearest Matching		Optimal Matching		Full Matching		Subclassification Matching
Covariate	Before	After	Before	After	Before	After	Before	After	Before	After
Distance to road (m)	−0.8182	−0.1277	−0.8192	−0.0706	−0.8182	−0.1244	−0.8182	−0.0818	−0.8182	−0.2399
Population density (people/km2)	−0.3273	−0.0814	−0.3286	0.1197	−0.3273	0.0445	−0.3273	0.0778	−0.3273	0.034
Elevation (m)	1.5725	0.1623	1.575	−0.0233	1.5725	0.1465	1.5725	−0.007	1.5725	0.1448
Slope (°)	0.8536	0.076	0.8564	0.0364	0.8536	0.0809	0.8536	0.0894	0.8536	0.2056
Precipitation (mm)	1.7338	0.0913	1.7341	−0.0007	1.7338	0.1693	1.7338	0.012	1.7338	0.1603
b. 2015
	1:1 Nearest Matching		1:2 Nearest Matching		Optimal Matching		Full Matching		Subclassification Matching
Covariate	Before	After	Before	After	Before	After	Before	After	Before	After
Distance to road (m)	−0.8182	−0.1077	−0.8192	−0.0766	−0.8182	−0.0513	−0.8182	0.0169	−0.8182	−0.1899
Population density (people/km2)	−0.3568	−0.0299	−0.3581	0.0999	−0.3568	−0.0331	−0.3568	0.0519	−0.3568	0.03
Elevation (m)	1.5725	0.2525	1.575	−0.0089	1.5725	0.2583	1.5725	0.0635	1.5725	0.166
Slope (°)	0.8536	0.0243	0.8564	0.0043	0.8536	0.0367	0.8536	0.075	0.8536	0.1972
Precipitation (mm)	1.9335	0.2925	1.9363	−0.0125	1.9335	0.3171	1.9335	0.0624	1.9335	0.2203
c. 2020
	1:1 Nearest Matching		1:2 Nearest Matching		Optimal Matching		Full Matching		Subclassification Matching
Covariate	Before	After	Before	After	Before	After	Before	After	Before	After
Distance to road (m)	0.2891	−0.1289	0.2902	−0.0849	0.2891	−0.1367	0.2891	−0.2731	0.2891	−0.1395
Population density (people/km2)	−0.3486	0.0862	−0.3499	0.083	−0.3486	0.1304	−0.3486	0.0761	−0.3486	0.0313
Elevation (m)	1.6004	0.0846	1.6029	−0.0117	1.6004	0.1058	1.6004	−0.0645	1.6004	0.1503
Slope (°)	0.9046	−0.0244	0.9075	0.0375	0.9046	−0.0379	0.9046	−0.0204	0.9046	0.0849
Precipitation (mm)	2.0914	0.1039	2.0942	−0.0038	2.0914	0.104	2.0914	−0.031	2.0914	0.2519

), confirming that these methods effectively reduced imbalance [57]. Based on these results, we finally selected 1:2 nearest-neighbor matching without replacement and set a caliper width of 0.2 in this study, which achieved relatively low SMDs while retaining as many samples as possible, and thus demonstrated the most robust performance [30]. Estimation was conducted using the MatchIt package (version 4.7.2) in R (version 4.5.0), with propensity scores generated through a logistic regression model. We also divided the buffer zone into five 10 km bands extending from the park boundary to further assess variations in conservation effects (

Table A2. Standard mean difference (SMD) of covariates before and after 1:2 nearest-neighbor matching.

a. 2010
	Total Sample		0–10 km		10–20 km		20–30 km		30–40 km		40–50 km
Covariate	Before	After	Before	After	Before	After	Before	After	Before	After	Before	After
Distance to road (m)	−0.8192	−0.0706	−0.2759	−0.0357	−0.9789	−0.3123	−0.632	−0.5859	−0.6758	−0.4166	−1.1445	−0.271
Population density (people/km2)	−0.3286	0.1197	−0.1706	0.147	−0.3839	0.184	0.261	0.11	−0.9123	0.1974	−0.2728	0.1023
Elevation (m)	1.575	−0.0233	1.2523	0.0541	1.5938	−0.0257	1.7322	−0.1192	1.6958	−0.1004	1.4876	−0.064
Slope (°)	0.8564	0.0364	0.703	0.0069	0.9053	0.1549	1.1079	−0.0835	0.9671	−0.0308	0.6415	−0.031
Precipitation (mm)	1.7341	−0.0007	1.2808	0.052	1.6407	−0.1203	1.8138	−0.0508	1.8653	−0.1461	1.7898	−0.2317
b. 2015
	Total Sample		0–10 km		10–20 km		20–30 km		30–40 km		40–50 km
Covariate	Before	After	Before	After	Before	After	Before	After	Before	After	Before	After
Distance to road (m)	−0.8192	−0.0766	−0.2759	−0.0875	−0.9789	−0.5116	−0.632	−0.8694	−0.6758	−0.8904	−1.1445	−0.2963
Population density (people/km2)	−0.3581	0.0999	−0.2083	0.1141	−0.4176	0.0648	0.2536	0.111	−0.9543	0.3587	−0.3019	0.0935
Elevation (m)	1.575	−0.0089	1.2523	0.0618	1.5938	−0.0294	1.7322	−0.2968	1.6958	−0.206	1.4876	−0.087
Slope (°)	0.8564	0.0043	0.703	0.0452	0.9053	0.0368	1.1079	−0.1367	0.9671	−0.1526	0.6415	−0.0494
Precipitation (mm)	1.9363	−0.0125	1.3983	0.0567	1.8383	−0.062	2.0502	−0.1853	2.1026	−0.1024	1.9702	−0.1242
c. 2020
	Total Sample		0–10 km		10–20 km		20–30 km		30–40 km		40–50 km
Covariate	Before	After	Before	After	Before	After	Before	After	Before	After	Before	After
Distance to road (m)	0.2902	−0.0849	0.4482	−0.1012	0.4011	−0.2829	0.2751	−0.4227	0.1592	−0.3439	0.2018	−0.0886
Population density (people/km2)	−0.3499	0.083	−0.2156	0.0907	−0.4199	0.1196	0.2606	0.0905	−0.9669	0.1948	−0.3076	0.0507
Elevation (m)	1.6029	−0.0117	1.2523	0.0428	1.5938	0.0842	1.7322	−0.0222	1.6958	0.0551	1.4877	0.0249
Slope (°)	0.9075	0.0375	0.703	0.0646	0.9053	−0.108	1.1079	−0.1433	0.9671	0.0075	0.6417	0.0168
Precipitation (mm)	2.0942	−0.0038	1.3732	0.0508	1.7769	0.1037	1.9426	−0.0349	1.9558	0.087	1.8057	0.0357

).

We defined the causal effects of conservation based on the “potential outcomes model” [58]. In this article, we evaluated the average treatment effect on the treated group (ATT) to reflect the conservation effectiveness with the following equations [57]:

$$\tau \left(X_i\right)=E\left[Y\left(0\right)-Y\left(1\right)|X_i\right]$$

(5)

$$ATT=E\left[\tau \left(X_i\right)|{NP}_i=1\right]$$

(6)

$\tau \left(X_i\right)$ represents the conditional average treatment effect, ${NP}_i$ represents the dummy variable, ${NP}_i=1$ represents the treated group (inside the national park), and $X_i$ represents the confounding variables.

2.3.3. Stratification-Multilevel Method for Heterogeneity Analysis

Individuals differ in their background and in how they react to interventions. Treatment results can also change depending on the chance of receiving treatment [59]. Recent research in conservation planning increasingly highlights the importance of considering economic constraints and prioritizing ecological objectives within a limited financial budget [60]. Investigating treatment effect heterogeneity offers critical insights into how conservation resources can be optimally allocated across protected areas, especially in diverse ecological and socio-economic contexts. This study employed a stratified multilevel approach to assess heterogeneity in treatment effects. The process involved three steps: (1) calculate propensity scores for all units using logistic regression on observed covariates (see Section 2.3.2); (2) divide the sample into five balanced strata of propensity scores, ensuring no meaningful differences in covariates or scores between treated and control groups; and (3) estimate treatment effects within each stratum based on its matched scores.

3. Results

3.1. Spatial-Temporal Patterns of Ecosystem Health

According to the results (Figure 3), the most notable growth was found in ES, particularly within the national park and the area directly to its north. By 2020, 59.2% of the study area had attained high levels of ES, and over 66% of the pixels exhibited positive growth, which correlated strongly with the expansion of eco-tourism in HNP and surrounding regions, as well as the implementation of the ‘closed rotation’ management system [16].

In terms of EV, the median values in 2010, 2015, and 2020 were 0.593, 0.678, and 0.655, respectively, showing an inverted “U-shaped” trend. Since 2013, Huangshan’s Degraded Forest Restoration Project has facilitated the transition of young forests into a rapid growth period, contributing to a 62.9% enhancement in EV. In 2020, EV remained close to the previous level but showed a slight decline. This was mainly due to the unprecedented heavy rainfall in July, which triggered severe flooding in the region and further reduced the total EV [61,62].

The study area exhibited a remarkably stable state (54.3%) of ER, with more than 70% of the pixels remaining at high resilience levels throughout the observation period. The mean ER values remained stable at 0.897, 0.895, and 0.896 during the study period, showing no statistically significant temporal variation. A total of 15.8% of the grid cells exhibited a declining trend, primarily observed in the basin area south of the national park, which could be largely attributed to anthropogenic activities from the urban areas.

From 2010 to 2020, the ecosystem health of HNP and its surrounding areas showed a gradual improvement, with the median increasing from 0.402 in 2010 to 0.439 in 2015 and then to 0.482 in 2020 (Figure 4a). The predominant EH grade shifted from moderate to high, and 69.5% of pixels showed significant enhancement over the decade-long study period. Overall, the spatial distribution of EH demonstrates superior conditions inside the national park compared to surrounding areas, with low-value zones predominantly occurring in urban regions of the northwest and southeast sectors.

Moreover, we identified a significant negative correlation between EH and the distance to the national park boundary (Figure 4b). From 2010 to 2020, interior park areas furthest from the boundaries maintained the highest EH status. EH decreased sharply from core protected areas to park boundaries, with the trend persisting in the 10 km buffer outside the national park. These findings provide clear evidence that national parks effectively preserve ecosystem health while generating positive spillover effects to surrounding landscapes.

3.2. Effectiveness of the National Park on Promoting EH

There was a mismatch in covariates between the national park and nearby unprotected lands. Compared to surrounding areas, the national park had a lower initial population density, as well as a higher initial elevation, slope, and annual precipitation (

Table A3. Mean covariate values of treated and control groups.

a. 2010
		Total Sample		0–10 km		10–20 km		20–30 km		30–40 km		40–50 km
Covariate		Treated	Control	Treated	Control	Treated	Control	Treated	Control	Treated	Control	Treated	Control
Distance to road (m)	before	2299.2998	3750.0526	2298.843	2788.2403	2298.843	4034.9366	2298.843	3419.8323	2298.843	3497.3887	2298.843	4328.7817
	after	2296.8441	2421.7849	2304.6852	2367.9645	2353.2238	2907.049	2371.4306	3410.6341	2438.5972	3177.4503	2340.1089	2820.7544
Population density (people/km2)	before	39.8905	75.6599	39.9759	58.5697	39.9759	81.8281	39.9759	11.5188	39.9759	139.4278	39.9759	69.7117
	after	39.6203	26.5857	41.068	25.0437	46.4267	26.3709	39.2163	27.2241	48.2512	26.7321	43.1658	32.0079
Elevation (m)	before	814.6774	344.6342	814.6798	440.3627	814.6798	338.3007	814.6798	296.9489	814.6798	307.821	814.6798	370.0524
	after	722.9438	729.9082	715.0471	698.8864	561.4861	569.1784	534.8763	570.5133	623.261	653.2831	659.9815	679.1186
Slope (°)	before	28.6589	20.4584	28.6397	21.9028	28.6397	19.9634	28.6397	18.0217	28.6397	19.3712	28.6397	22.492
	after	27.7201	27.3711	27.1925	27.1263	26.5667	25.0825	25.0202	25.8207	26.4656	26.761	27.2732	27.5701
Precipitation (mm)	before	2158.6183	1917.4703	2158.8503	1980.5435	2158.8503	1930.4489	2158.8503	1906.347	2158.8503	1899.1792	2158.8503	1909.682
	after	2119.5538	2119.6498	2115.3079	2108.0651	2054.3095	2071.0627	2047.8952	2054.9657	2087.6172	2107.9552	2094.4999	2126.7496
b. 2015
		Total Sample		0–10 km		10–20 km		20–30 km		30–40 km		40–50 km
Covariate		Treated	Control	Treated	Control	Treated	Control	Treated	Control	Treated	Control	Treated	Control
Distance to road (m)	before	2299.2998	3750.0526	2298.843	2788.2403	2298.843	4034.9366	2298.843	3419.8323	2298.843	3497.3887	2298.843	4328.7817
	after	2343.6505	2479.3505	2260.8373	2415.9495	2460.4962	3367.8896	2727.1873	4269.2183	2949.3417	4528.5073	2494.007	3019.5325
Population density (people/km2)	before	36.4278	75.0241	36.5066	58.9943	36.5066	81.576	36.5066	9.133	36.5066	139.5122	36.5066	69.0929
	after	39.0544	28.2871	38.4965	26.1793	29.8832	22.8872	34.8755	22.8953	58.71	19.9926	46.2367	36.1451
Elevation (m)	before	814.6774	344.6342	814.6798	440.3627	814.6798	338.3007	814.6798	296.9489	814.6798	307.821	814.6798	370.0524
	after	726.5854	729.2512	723.4696	704.9945	559.0807	567.8628	511.4322	600.1312	630.3683	691.9307	603.027	629.0275
Slope (°)	before	28.6589	20.4584	28.6397	21.9028	28.6397	19.9634	28.6397	18.0217	28.6397	19.3712	28.6397	22.492
	after	27.3952	27.3541	27.6572	27.2243	26.3665	26.0135	25.0673	26.3775	25.2971	26.7592	26.3219	26.7954
Precipitation (mm)	before	2429.0029	2125.386	2429.0371	2209.4455	2429.0371	2140.3498	2429.0371	2107.0823	2429.0371	2098.8471	2429.0371	2119.6327
	after	2376.1734	2378.1334	2372.7792	2363.8683	2285.4191	2295.1616	2267.1484	2296.2406	2315.7657	2331.8468	2312.6804	2332.1895
c. 2020
		Total Sample		0–10 km		10–20 km		20–30 km		30–40 km		40–50 km
Covariate		Treated	Control	Treated	Control	Treated	Control	Treated	Control	Treated	Control	Treated	Control
Distance to road (m)	before	2137.7774	1720.1455	2136.8252	1490.9546	2136.8252	1558.734	2136.8252	1740.3323	2136.8252	1907.4124	2136.8252	1846.01
	after	2162.8421	2284.9909	2072.9991	2218.896	2173.1071	2580.8344	2338.7935	2948.025	2276.1901	2771.8036	2094.7608	2222.5002
Population density (people/km2)	before	35.9829	73.0048	36.0603	58.9012	36.0603	80.5498	36.0603	8.4471	36.0603	138.518	36.0603	68.6497
	after	36.7455	27.9584	35.2438	25.6339	33.2056	20.5373	31.8279	22.2396	38.2269	17.5804	40.6419	35.2743
Elevation (m)	before	814.6774	336.3074	814.6798	440.3627	814.6798	338.3007	814.6798	296.9489	814.6798	307.821	814.6798	370.0258
	after	757.3263	760.8066	717.398	704.5992	633.5675	608.3949	594.2765	600.8972	730.7725	714.2959	703.4744	696.0442
Slope (°)	before	28.6589	19.969	28.6397	21.9028	28.6397	19.9634	28.6397	18.0217	28.6397	19.3712	28.6397	22.4898
	after	28.2641	27.9047	27.4034	26.784	25.7991	26.8338	26.13	27.5036	27.8387	27.7665	27.7925	27.6312
Precipitation (mm)	before	2120.3072	1827.7762	2120.3783	1928.273	2120.3783	1871.8062	2120.3783	1848.6164	2120.3783	1846.7676	2120.3783	1867.7727
	after	2089.446	2089.9827	2068.3312	2061.2286	2028.807	2014.2997	2006.2251	2011.1011	2081.6965	2069.5315	2060.961	2055.9651

). These factors show that baseline gaps existed in ecosystem health across the treatment and control groups (

Table A4. Mean values of ecosystem health (EH), ecosystem service (ES), ecosystem resilience (ER), and ecosystem vigor (EV) for treated and control groups.

a. 2010
	EH	ES	ER	EV
Treated	0.4050037	0.10959544	0.9879207	0.6190458
Control	0.3819352	0.09890886	0.9581496	0.6017086
b. 2015
	EH	ES	ER	EV
Treated	0.4321984	0.1213692	0.9851929	0.6829016
Control	0.4171342	0.1131893	0.9644158	0.6810368
c. 2020
	EH	ES	ER	EV
Treated	0.4812389	0.1709042	0.9865474	0.6672542
Control	0.4508396	0.1525037	0.9549505	0.6468838

). To address this, we applied propensity score matching to choose nearby sites with comparable conditions to the park, which served as control samples. This approach effectively reduced covariate differences between the park and the surrounding areas.

According to the propensity score matching model, the ATT results for EH were significantly positive for three years, indicating that the national park had a positive conservation impact on improving EH (Figure 5). In 2010, 2015, and 2020, the average treatment effects on the treated group were 0.022, 0.014, and 0.029. Therefore, in comparison to the unprotected situation, the overall EH values were promoted by 5.7%, 3.4%, and 6.5%, respectively.

Specifically, the national park performed best in enhancing ES. From 2010 to 2020, conservation effects promoted ES in the national park by 10%, 7%, and 11.8% through better water restoration, soil conservation, air quality, and cultural or recreational services. Similarly, the national park contributed steadily to ER improvement, with ATT values exceeding those of non-protected areas by over 0.02 in all three years. However, EV is a notable exception. In 2015, the region’s overall EV peaked, with both inside and outside the park showing high vigor, which made the park’s conservation advantage less apparent. Yet, the gap widened again in 2020 (0.018), mainly because flooding reduced EV across the region, with low-altitude areas outside the park suffering the most. In contrast, the ecosystems inside the park were more resilient, which increased the difference between inside and outside, as shown by the higher ATT values.

We also found that the spillover effects of the national park vary with distance. By comparing samples from each 10 km buffer zone with those in the national park (Figure 6), we discovered that most of the ATT values of ES, ER, and EH within 30 km buffer zones showed a gradual upward trend with the increasing distance from the park border. Since ATT reflects the difference between treated and control groups, increasing ATT values indicate that ecological conditions in buffer zones diverge more from the park as distance grows [43], showing the typical characteristics of a positive spillover effect.

To be specific, in terms of ER, the 20–30 km buffer zone maintained a relatively high ATT value (0.046, 0.070, 0.063) over the past decade, while the ATT value of the 0–10 km buffer zone remained stable at 0.020–0.021, always lower than other areas, highlighting that the conservation measures of the national park had provided continuous support for ecosystem resilience in neighboring areas. As for ES, the impact of the national park also showed an obvious distance attenuation pattern within the 0–30 km buffer zone, which means that gateway communities closer to the national park will have better access to natural resources and ecotourism development opportunities. Regarding the EV, both 2010 and 2015 witnessed a slight leakage effect, as areas closer to the park boundary showed higher ATT values. It suggested that protection inside the park limited disturbances but also shifted some human pressures, such as logging or farming [63,64], to areas just outside the boundary. As a result, vegetation vigor near the park edge declined, even though overall EV in the region, especially in 2015, remained high. By 2020, however, the pattern changed. Extreme weather reduced EV outside the park more severely, while ecosystems inside remained more resilient, leading ATT values to follow the broader trend of ecosystem health.

3.3. Heterogeneity in the Conservation Effects of the National Park

Although our results confirm that the national park contributes to enhancing ecosystem health, the extent of these benefits may differ depending on the starting ecological conditions. To examine this variation, we used a stratification–multilevel method to evaluate how the impacts of treatment shift across different propensity score groups and covariate patterns.

For areas with better initial ecosystem health conditions, the ATT estimates suggest that the national park substantially enhances ecosystem health. As shown by the right side of Figure 7a, the ATT values steadily increased as the propensity score rose above 0.6, indicating stronger positive effects in these regions. Areas with low propensity scores (<0.2) also showed relatively high and statistically significant treatment effects. However, as ecological conditions improved, the conservation effect declined. We summarize this pattern as evidence of diminishing marginal returns in conservation effectiveness, meaning degraded areas bring higher gains from extra measures. For regions with a medium level of initial ecological condition (propensity scores between 0.2 and 0.6), the impact of the national park appears to be more complex. As shown in the middle range of Figure 7a, ATT values exhibited a declining trend and became statistically non-significant. This pattern reflected a transitional space where conservation impacts may be mediated by factors such as mixed land use, fragmented habitats, or competing socioeconomic pressures. These results highlighted the importance of considering ecological and social context when evaluating the effectiveness of protected areas across a spectrum of baseline conditions.

According to the heterogeneity analysis of the covariates (Figure 7b), we discovered that conservation effects on EH were more obvious in areas with lower elevation, less annual precipitation, and smaller slopes. This also reflects the principle of diminishing marginal benefits, as these areas face greater human pressure [57] and therefore have more room for ecological improvement. The correlation between population density and ATT exhibits a nonlinear pattern. In natural areas with low population density, the effectiveness of protecting EH decreases sharply or even becomes ineffective, suggesting that the area should be strictly preserved in its natural state because the environment is extremely sensitive to even the smallest human activity. As population density increased, ATT values rebounded briefly but then showed a downward trend, indicating that there was a “threshold” for management measures, excessive tourism activities and population aggregation in local communities had a negative impact on conservation efforts. The ATT values did not change significantly with the distance from the road, but the confidence interval is wider in the area close to the road, indicating that the estimation of ATT is interfered by complex human factors such as fluctuations in the number of tourists and high uncertainty.

4. Discussion

4.1. Characteristics of Ecosystem Health Changes in HNP and Its Surrounding Areas

Based on the VSR (Vigor–Service–Resilience) framework, this study comprehensively considered the intrinsic health of ecosystems while innovatively incorporating human well-being into the evaluation criteria [23]. The overall ecosystem health in Huangshan National Park and its surrounding areas showed a positive trend between 2010 and 2020. This outcome is consistent with the conclusions reported by Ma et al. (2024) [16]. The past decade has witnessed a significant improvement in ecosystem health in the region, driven by the implementation of ecological compensation policies and the expansion of the leisure and education industries [32,65]. In addition, as a dynamic indicator of ecosystem productivity, ecosystem vigor is particularly sensitive to disruptions [48]. Our results align with previous findings showing that extreme rainfall in the southern Anhui mountainous region reduces photosynthetic activity [62], and the flooding in the summer of 2020 further suppressed productivity [61], underscoring the importance of timely and appropriate forest management.

4.2. The Remarkable Contributions of the HNP on Promoting Ecosystem Health

Causal inference analysis confirms the effectiveness of the national park in enhancing ecosystem health. By using the PSM model, this study overcomes the shortcomings of previous spatiotemporal analyses based on GIS and remote sensing, minimizing errors and baseline bias caused by confounding environmental factors [7,57]. It is worth noting that the national park performed best in enhancing ecosystem services. The results support existing literature that national parks generate economic benefits by enhancing local habitat quality and attracting ecotourism, recreation, and environmental education opportunities [43], thereby creating a positive feedback loop for conservation investments [66].

To better explore the spatial spread of conservation effects in surrounding areas, five 10 km wide buffer zones outside the HNP were generated. The majority of conservation efforts on ES, ER, and the overall EH within 30 km buffer zones showed a gradual downward trend with the increasing distance from the park border. This pattern supports the view of Ament and Cumming (2016) [67] that national parks generate positive spillover effects by encouraging nearby land users to maintain the ecological environment for nature-based ecotourism. However, both 2010 and 2015 witnessed a slight leakage effect in EV, suggesting that resource demands may have shifted to nearby areas and caused deforestation and vegetation loss [63,68]. This speculative interpretation needs to be verified in future studies using more detailed in situ data. It is also important to note that spatial spillover effects are scale-dependent [68]. Previous studies constantly applied a single buffer radius of 10 or 20 km [64,69], ignoring that the extent of spillover possibly varied with sizes and distance from the protected areas [63]. Our findings confirm that the influence of national parks on ecosystem health can reach 30 km or more beyond their borders. This suggests that areas closest to national parks, especially gateway communities, should make use of their stronger ecological resilience and better access to natural resources and ecotourism opportunities to improve local well-being [70,71].

4.3. Heterogeneous Conservation Effects Across Ecological Baselines

Protected areas need to balance protection efficiency, costs, and benefits [72]. Although more studies are now paying attention to the outcomes of conservation policies, few have linked this evidence to spatially explicit protection priorities that could guide region-based conservation planning [30]. In order to develop more equitable and precise zoning management strategies, this study innovatively employed a stratification–multilevel method to analyze the effectiveness of conservation measures across different socio-ecological conditions. This method addresses a key research gap by moving beyond single-dimension ecosystem indicators and helps ensure that limited resources are allocated for the greatest ecological gains [60]. The heterogeneity analysis showed that conservation benefits were most pronounced in areas with lower elevations, gentler slopes, and less precipitation, a pattern consistent with previous research [30,73]. At the same time, some scholars argue that rigorous management measures enhance the performance of protected areas [74], while others present contrary findings [75]. Our results support the conclusions in the literature that areas with better ecological conditions experience stronger conservation effects and should be prioritized for protection [76]. In addition, we also observed a pattern of diminishing marginal returns that regions with poorer ecological baselines often showed greater improvements after protection, but these gains decreased as conditions improved [30]. This can be explained by the weaker capacity of degraded areas to provide ecosystem services and self-regulate, which makes the benefits of protection more evident. To maximize the effectiveness of conservation interventions, management should integrate both area-based and outcome-based approaches [24]. This is especially important in regions where strict protection is difficult to enforce [77]. Overall, these findings offer planners and decision-makers a spatially explicit, results-based approach to account for heterogeneity in conservation effectiveness, which is especially valuable when facing trade-offs or budget constraints [78].

4.4. Managerial Implications

Based on the previously mentioned analysis, the following management suggestions are provided to improve ecosystem health conservation in national parks: (1) Incorporating an ecosystem services approach can strengthen the synergy between biodiversity conservation and ecosystem service provision in planning [21,79]. It is recommended that the management committee strengthen ecological compensation mechanisms while enhancing disaster-prevention and forest-maintenance measures. This should include sustainable logging regulations, reforestation programs, and continuous ecological monitoring. Such an integrated approach can sustain long-term improvements in ecosystem services and resilience while reducing the vulnerability of ecosystem vigor to climate disturbances. (2) Given the heterogeneity of conservation costs and benefits across different landscape types, it is necessary to implement a zoning management strategy. For wilderness areas with better ecological conditions, strict protection should be maintained. At the same time, it is important to recognize the diminishing marginal benefits of protection and give priority to areas with lower elevation, less rainfall, and gentler slopes, where conservation gains are higher. Such targeted interventions not only safeguard ecosystems, but also ensure efficient resource allocation and maximize conservation effectiveness [8]. (3) Recognizing the significant spillover effects of the national park on surrounding buffer zones, future management should prioritize conserving and restoring critical connectivity areas [80]. At the same time, conservation efforts need to be integrated with local economic development in ways that align with both community interests and park objectives [81]. In particular, promoting eco-friendly enterprises such as ecotourism in buffer zones can strengthen ecosystem conservation, support species protection, and improve local livelihoods.

4.5. Limitations and Future Research

It is important to acknowledge certain limitations and uncertainties of this study. Due to data availability, the analysis relied on information from a specific period. Future work should aim to compile more comprehensive datasets, including extended hydrological and meteorological records, detailed land use change data, and long-term, high-resolution remote sensing images. Such data would allow for deeper, multi-scale analyses of ecosystem health dynamics and patterns [10]. In addition, although this study used counterfactual analysis to identify both positive spillover and slight leakage effects in areas surrounding the national park, future research should qualify the heterogeneity of these effects across regions and explore the natural and social factors driving them. This would provide more nuanced insights for the sustainable use of areas adjacent to national parks, especially for the gateway communities. Despite these limitations, the study offers a useful framework for evaluating ecosystem health and developing cost-effective spatial conservation strategies. Future research could apply this framework to other national parks in China for comparison and validation, or conduct scenario simulations to explore how ecosystem health might change under different management or land use interventions, thereby maximizing the benefits of conservation measures [79,82].

5. Conclusions

This study examined the conservation effectiveness of Huangshan National Park (HNP) in promoting ecosystem health between 2010 and 2020 by integrating the Vigor–Service–Resilience (VSR) framework with causal inference models. Results showed that ecosystem health improved significantly across the park and its surrounding areas, with 69.5% of the study area experiencing positive changes. Compared with matched unprotected sites, HNP enhanced overall ecosystem health by 5.7%, 3.4%, and 6.5% in 2010, 2015, and 2020, respectively. Positive spillover effects were detected up to 30 km beyond park boundaries, underscoring the broader regional benefits of conservation. Heterogeneity analysis revealed that conservation outcomes were not uniform. Areas with lower elevation, gentler slopes, and less precipitation benefited most, while diminishing marginal returns were observed as ecological conditions improved. These findings highlight the need for precision-oriented conservation strategies that balance strict protection in ecologically intact areas with outcome-based interventions where ecological gains are greatest. This study advances previous work by applying a robust counterfactual framework to establish clear causal links between national park policies and ecosystem health outcomes, offering valuable insights for achieving the 30 × 30 conservation target. In future work, greater attention should be given to long-term data monitoring, zoning strategies, and gateway community management to maximize the effectiveness of conservation interventions.