Dynamic evolution of alpine grassland social-ecological resilience and its characteristics based on the “components-attributes-capacities” framework: a case study of the Qinghai-Tibet Plateau

1 Introduction

Ecological conservation of the Qinghai-Tibet Plateau has long been recognized as a strategic priority at the national level. Since the 18th National Congress of the Communist Party of China, the Chinese government has incorporated ecological civilization development into the overall framework of the “Five-Sphere Integrated Plan,” emphasizing that “lucid waters and lush mountains are invaluable assets” and highlighting that “protecting the ecological environment of the Qinghai-Tibet Plateau represents the greatest contribution to the survival and development of the Chinese nation.” In August 2020, President Xi Jinping further emphasized at the Seventh Central Tibet Work Symposium that efforts should be made to transform the Qinghai-Tibet Plateau into a leading ecological civilization region of national and global significance. In 2023, the implementation of the Qinghai-Tibet Plateau Ecological Conservation Law of the People’s Republic of China marked a transition toward a new stage of legalized and systematic ecological governance. Furthermore, the “15th Five-Year Plan” (2026-2030) identified the Qinghai-Tibet Plateau ecological barrier zone as the foremost component of China’s national ecological security barrier system and explicitly called for the further advancement of ecological conservation efforts on the plateau. These strategic initiatives and institutional safeguards underscore the irreplaceable role of the Qinghai-Tibet Plateau as a critical national ecological security barrier. Known as the “Roof of the World” and the “Asian Water Tower,” the Qinghai-Tibet Plateau covers an area of approximately 2.5 million km², with an average elevation ranging from 4000 to 5000 m. It represents one of the most concentrated regions of high-altitude biodiversity worldwide (Dong, 2023; Zhu et al., 2023; Wang et al., 2026). As a vital ecological security barrier for China and Asia, the alpine grassland ecosystems of the Qinghai-Tibet Plateau provide essential ecosystem services, including carbon sequestration, water conservation, and biodiversity maintenance. Moreover, these ecosystems play an irreplaceable role in regulating regional climate patterns and sustaining water security in surrounding countries (Feng et al., 2024; Liu et al., 2025). Consequently, the plateau constitutes a key region for implementing national ecological security strategies and advancing the vision of harmonious coexistence between humans and nature.

However, under the combined pressures of accelerating climate change and increasing anthropogenic disturbances, alpine grassland ecosystems in this region are experiencing varying degrees of degradation risk. Consequently, maintaining structural stability and sustaining ecosystem service functions have become frontier issues in global change research and regional sustainable development. The alpine grasslands of the Qinghai-Tibet Plateau cover approximately 1.28×10⁸hm², accounting for 33.2% of China’s total grassland area (Wei et al., 2023; Jiang et al., 2025; He et al., 2026). Their ecological functions are not only closely associated with the livelihoods of local pastoral communities and regional economic development but also exert profound influences on climate regulation and hydrological cycles across East Asia and beyond. Unlike temperate grassland systems such as those in Inner Mongolia, alpine grasslands rely on fragile permafrost-vegetation coupled systems, characterized by short growing seasons, extremely prolonged self-recovery periods, delayed ecological compensation capacity, and highly vulnerable livestock-grass balance mechanisms (Wang et al., 2023; Wang et al., 2024; Huang et al., 2026). In recent decades, climate warming-induced permafrost degradation, altered precipitation regimes, and intensified anthropogenic disturbances, including overgrazing and engineering activities, have jointly exacerbated grassland degradation. Typical manifestations include the expansion of degraded “black soil patches,” proliferation of toxic and invasive weeds, frequent outbreaks of rodent and insect pests, and intensified soil erosion, resulting in ecological pressures that substantially exceed those of low-elevation grasslands at comparable latitudes (Zhou et al., 2023; Wu et al., 2024; Ji et al., 2025). Meanwhile, deficiencies in cross-regional ecological compensation mechanisms, together with administrative fragmentation and challenges in coordinating stakeholder interests, have further constrained the implementation of integrated ecological governance (Deng et al., 2026). Facing such a complex landscape characterized by multiple interacting disturbances, scientifically assessing the resilience of alpine grasslands, elucidating their spatiotemporal evolutionary patterns, and identifying underlying driving mechanisms have become critical scientific challenges for safeguarding ecological security on the Qinghai-Tibet Plateau and achieving coordinated high-quality development of pastoral economies and ecosystems.

The concept of resilience originated from the field of mechanical engineering and was subsequently introduced into ecology by Holling to describe the capacity of ecosystems to recover from disturbances while maintaining their functional structures (Holling, 1973). With the advancement of resilience research, the concept has gradually expanded into the analytical framework of coupled social-ecological systems, becoming a fundamental theoretical tool for understanding the abilities of complex systems to absorb external shocks, adapt to changing conditions, and undergo transformative change (Scheffer et al., 2001; Ostrom, 2004; Walker et al., 2004).In the fields of ecology and regional sustainable development, resilience is no longer interpreted merely as the capacity of a system to return to its original state after disturbance. Instead, increasing attention has been directed toward the potential of systems to achieve adaptive transformation through self-adjustment, learning, and innovation in response to uncertainty. For alpine grassland systems on the Qinghai-Tibet Plateau, resilience encompasses not only the recovery dynamics of ecological attributes, such as vegetation coverage and ecosystem productivity, but also the capacity of pastoral economies to withstand external shocks, the robustness of social public service systems, and the effectiveness of policy interventions (Li et al., 1997; Shao et al., 2006; Li et al., 2024). Therefore, developing a resilience assessment framework that integrates ecological, economic, and social subsystems while incorporating three dimensions of capacity-resistance, recovery, and adaptation is of substantial theoretical significance and practical relevance. In this study, “alpine grassland resilience” specifically refers to the resilience of the alpine grassland social-ecological system, rather than the resilience of the grassland ecosystem alone. This distinction is important because alpine grassland resilience is jointly shaped by ecological conditions and socioeconomic contexts at the prefecture level. The economic and social indicators included in this study do not directly measure grassland ecological status; instead, they reflect the broader regional socioeconomic environment that supports, constrains, and regulates the functioning of the alpine grassland social-ecological system. Prefecture-level socioeconomic conditions may influence grassland resilience through multiple pathways, including livelihood diversification, ecological governance investment, infrastructure support, public service provision, policy implementation capacity, and pastoralists’ adaptive behavior. Therefore, the study object is consistently defined as the resilience of the alpine grassland social-ecological system.

In recent years, extensive research has been conducted on grassland ecosystem resilience, covering various aspects including vegetation dynamics monitoring, responses to climate change, optimization of grazing management, and evaluation of ecological compensation policies (Wang et al., 2025; Zhao et al., 2025; Augustine et al., 2026; Liu et al., 2026). Liu et al. (2026) employing remote sensing and machine learning approaches, demonstrated that precipitation variability was the dominant factor driving the decline in grassland resilience across China during 2001-2021, with a substantially greater contribution than grazing activities; notably, resilience degradation was particularly pronounced on the Qinghai-Tibet Plateau (Zhang Y. et al., 2025). Zhao et al. (2025) revealed that extreme precipitation events explained 65.68% and 75.94% of the variations in net primary productivity (NPP) and vegetation resilience (VR), respectively, across the Qinghai-Tibet Plateau, and identified a critical ecological threshold of approximately 2.4 for grazing pressure (Hu et al., 2026). Hu et al. (2026) based on data from 443 field sampling sites, found significant differences in the mechanisms driving ecosystem stability among alpine meadows, alpine steppes, and alpine deserts, with grazing intensity exerting a negative effect on alpine meadow stability (Liu et al., 2026). Liu et al. (2026) using the Bayanbulak grassland in the central Tianshan Mountains as a case study, further demonstrated that the implementation of China’s grassland ecological compensation policies did not establish a simple positive relationship between grassland “greening” and ecosystem stability enhancement (Cai et al., 2025). At the household level, Cai et al. (2025) found that the impacts of grassland ecological compensation policies on livelihood resilience varied significantly among pastoral households with different livestock production scales. Collectively, these studies have advanced the understanding of grassland resilience across multiple spatial and analytical scales (Lester and Fairweather, 2011). Cheng et al. (2025) constructed a cropland ecosystem-health assessment framework and used an XGBoost-SHAP approach to separate the contributions of climatic factors and human activities to ecosystem-health dynamics. Their study showed that climate change and human activities may exert heterogeneous and nonlinear effects across space and time, and that SHAP-based interpretation can effectively reveal the relative importance and interaction mechanisms of different drivers (Cheng et al., 2025). However, most existing studies have focused on individual ecological processes or specific regions, providing limited insights into the systematic and dynamic characteristics of resilience of the alpine grassland social-ecological system. This limitation is particularly evident on the Qinghai-Tibet Plateau, where extreme environmental conditions, difficulties in data acquisition, and pronounced internal heterogeneity pose substantial challenges for resilience assessment. Consequently, current approaches for quantifying alpine grassland resilience remain insufficiently developed, while the spatiotemporal evolution patterns, resilience capacity combination modes, and their underlying mechanisms influencing overall resilience remain poorly understood. Furthermore, previous studies have predominantly relied on static indicators or linear modelling approaches, which are insufficient for capturing the complex characteristics of resilience systems characterized by the coexistence of gradual transitions and abrupt shifts, as well as frequent nonlinear feedback processes. Therefore, analytical frameworks capable of addressing nonlinear relationships, multidimensional interactions, and dynamic evolutionary processes are urgently needed to advance the comprehensive assessment and adaptive governance of resilience of the alpine grassland social-ecological system.

To address these research gaps, this study focuses on 21 prefecture-level administrative units across the Qinghai-Tibet Plateau and integrates multiple analytical approaches, including the catastrophe progression method, three-dimensional kernel density estimation, Markov chains, and self-organizing feature mapping (SOFM) networks. A three-dimensional resilience assessment framework based on the progressive “Components-Attributes-Capacities” structure is developed to systematically investigate the spatiotemporal evolution patterns, resilience level transition dynamics, and capacity combination modes of resilience of the alpine grassland social-ecological system from 2000 to 2024. The objectives of this study are to elucidate the underlying mechanisms driving resilience of the alpine grassland social-ecological system evolution and to provide theoretical insights and empirical evidence for enhancing ecological security and promoting sustainable development in pastoral regions of the Qinghai-Tibet Plateau.

2 Theoretical framework and assessment system

This study examines prefecture-level alpine grassland social-ecological resilience capacity on the Qinghai-Tibet Plateau. The prefecture-level administrative unit is adopted as the unit of analysis because ecological conservation, pastoral development, public-service provision, ecological compensation, and regional governance are largely coordinated at this scale. Accordingly, the assessment does not measure resilience at the grassland-patch, household, or individual level. Rather, it characterizes the ecological, economic, social, and governance-related conditions that jointly support the capacity of alpine grassland social-ecological systems to cope with disturbances and long-term environmental change.

A “Components–Attributes–Capacities” framework was developed to establish the conceptual basis of the assessment. The components layer refers to the ecological, socioeconomic, and governance-related conditions that provide the material, environmental, and institutional basis for the functioning of the alpine grassland social-ecological system. These conditions include climate, vegetation, soil, pastoral production, livelihood resources, public services, infrastructure, and policy support (Henry and Ramirez-Marquez, 2017; Chu et al., 2023). The attributes layer comprises resistance, recovery, and adaptation. Resistance refers to the ability of a system to absorb or withstand disturbances while maintaining essential functions; recovery refers to the ability to restore or reorganize functions following stress or disturbance; and adaptation refers to the ability to adjust livelihood strategies, management practices, and institutional arrangements in response to changing environmental and socioeconomic conditions (Ge et al., 2024). The capacities layer is represented by ecological, economic, and social capacities, which jointly determine the overall resilience capacity of the alpine grassland social-ecological system (Xu and Guo, 2015; Li et al., 2025).

To ensure consistency between the conceptual framework and the empirical assessment, the indicator system is organized into a four-level operational hierarchy: overall resilience capacity–subsystem capacity–resilience attribute–proxy indicator (Table 1). Specifically, ecological, economic, and social subsystems constitute the three capacity domains; resistance, recovery, and adaptation represent the functional attributes within each domain; and the 29 indicators are used as proxy measures of the component conditions and enabling resources associated with these attributes. Thus, the ecological, economic, and social subsystems are not separate from the “Components–Attributes–Capacities” framework; rather, they operationalize the capacity layer of the framework.

Table 1. Evaluation indicator system for alpine grassland social-ecological system resilience on the Qinghai-Tibetan Plateau.

All indicators are interpreted as proxy measures of resilience-related capacities and enabling conditions at the prefecture level. They do not directly measure realized resistance, post-disturbance recovery trajectories, or adaptive outcomes following a specific disturbance event.

The ecological subsystem reflects the internal ecological condition and self-regulation capacity of alpine grasslands. Its indicators include vegetation coverage, net primary productivity (NPP), soil organic carbon content, grassland type equilibrium, and other variables related to ecosystem structure, productivity, and functional maintenance (Zhang et al., 2020; Li et al., 2024; Lloret et al., 2024). The economic subsystem represents the material and livelihood basis for maintaining and restoring system functioning. Indicators such as pastoral industrial diversity, per capita income, fiscal self-sufficiency, and infrastructure conditions reflect the capacity of local economies to buffer livelihood shocks, support post-disturbance recovery, and facilitate livelihood adjustment. The social subsystem characterizes governance, public-service provision, social protection, and collective adaptive capacity. Educational and healthcare resources, social security coverage, veterinary service accessibility, policy awareness, and related indicators represent the institutional and social conditions that enable communities to respond to environmental and socioeconomic pressures (Sun and Feng, 2013; Tofu et al., 2023; Li, 2026).

It should be emphasized that the indicators in Table 1 are interpreted as proxy measures of resilience-related capacities and enabling conditions, rather than direct observations of realized resistance, post-disturbance recovery trajectories, or adaptive outcomes following a specific disturbance event. For example, ecological compensation coverage and ecological-protection expenditure represent policy support for ecological management; social security coverage represents social protection capacity; and e-commerce adoption reflects the potential for livelihood diversification and market adaptation. These variables may strengthen resilience capacity, but they do not independently demonstrate that a system has already resisted, recovered from, or adapted to a particular disturbance. Therefore, the composite index developed in this study should be interpreted as a measure of prefecture-level social-ecological resilience capacity or resilience-related condition.

Based on this framework, 29 proxy indicators were selected to characterize resilience-related capacities and enabling conditions across the ecological, economic, and social subsystems at the prefecture level (Table 1). The catastrophe progression method was used to aggregate the indicators and estimate the composite resilience-capacity index. This method incorporates catastrophe models and normalization functions to characterize nonlinear changes in indicator performance while reducing the subjectivity associated with conventional weighting procedures (Zhu et al., 2013).

To examine the spatiotemporal evolution and structural differentiation of resilience capacity, this study further integrated three-dimensional kernel density estimation, Markov chain analysis, and self-organizing feature mapping (SOFM) clustering. Three-dimensional kernel density estimation was used to identify changes in the distribution, convergence, and polarization of resilience capacity over time. Markov chain analysis was applied to quantify state persistence and transition probabilities among different resilience-capacity levels, thereby revealing potential path dependence and convergence patterns. SOFM clustering was used to identify typical combinations of ecological, economic, and social resilience attributes without imposing predefined categories (Kohonen, 2001; Rousset et al., 2006). Together, these methods provide a dynamic and multidimensional perspective on the evolution, transition, and configuration of prefecture-level alpine grassland social-ecological resilience capacity.

3 Data sources and research methods

3.1 Study area and data sources

The study area comprises 21 prefecture-level administrative units on the Qinghai-Tibet Plateau where alpine grasslands are widely distributed and where relatively complete ecological and socioeconomic data are available. The selected units include Xining, Haidong, Haixi, Haibei, Hainan, Huangnan, Guoluo, and Yushu in Qinghai Province; Lasa, Linzhi, Shannan, Ali, Rikaze, Naqu, and Changdu in Tibet Autonomous Region; Ganzi and Aba in Sichuan Province; Gannan in Gansu Province; Diqing in Yunnan Province; and Kezhou, Kashi in Xinjiang Uygur Autonomous Region(Figure 1). The inclusion criteria were as follows: (1) the unit contains a substantial area of alpine grassland; (2) complete and continuously available socioeconomic statistics could be obtained for the study period; and (3) the unit has sufficient ecological and geospatial data for longitudinal analysis.

The Qinghai-Tibet Plateau, the world’s highest and most extensive alpine ecosystem, is characterized by an average elevation exceeding 4,000 m, low temperatures, hypoxia, and a short growing season. Its alpine grasslands depend on fragile permafrost-vegetation coupled systems, in which low biological productivity and slow nutrient cycling result in prolonged recovery following disturbance. These ecosystems also provide essential services, including water conservation, soil retention, biodiversity maintenance, and carbon storage. Under the combined effects of climate warming, permafrost degradation, grazing pressure, and localized engineering disturbances, alpine grasslands have experienced increasing ecological stress. Degraded “black soil patches,” the spread of toxic and undesirable plant species, and rodent and insect disturbances can further reduce vegetation cover and intensify soil erosion. Compared with grasslands at lower elevations, the livestock-grass balance in this region is particularly vulnerable because of limited forage production, strong climatic constraints, and slow ecological recovery. These environmental characteristics create substantial spatial heterogeneity in grassland resource conditions, disturbance regimes, and recovery capacity. Understanding the dynamic evolution of resilience in alpine grassland social-ecological systems is therefore important for maintaining ecological security and supporting the sustainable development of pastoral livelihoods.

The datasets used in this study include two major categories: (1) remote sensing and ecological monitoring data and (2) socioeconomic statistical data. The land-use and vegetation cover datasets were obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS) remote sensing products provided by NASA Earthdata (https://www.earthdata.nasa.gov). These datasets were used to derive the normalized difference vegetation index (NDVI) and a proxy for habitat richness. Net primary productivity (NPP) and soil organic carbon (SOC) content were obtained from the U.S. Geological Survey (USGS) Land Processes Distributed Active Archive Center (LP DAAC; https://lpdaac.usgs.gov), including the MOD17A3 product (Running et al., 2004) and global soil carbon pool datasets (Poggio et al., 2021), respectively. Grassland type equilibrium was calculated using landscape pattern indices based on land-cover data obtained from the Resources and Environmental Science Data Center of the Chinese Academy of Sciences (RESDC; https://www.resdc.cn). Socioeconomic statistical data were collected from multiple official statistical sources, including the China City Statistical Yearbook, China Animal Husbandry Statistical Yearbook, China Rural Statistical Yearbook, China Forestry and Grassland Statistical Yearbook, China Financial Yearbook, and China Educational Statistics Yearbook. Additional information was obtained from provincial statistical yearbooks, grassland monitoring reports, and fixed-point monitoring datasets for pastoral regions. Missing values were processed using the multiple imputation method. Specifically, multiple imputation by chained equations (MICE) was applied with m=5 imputations and 10 iterations per imputation, using a random seed of 12345.Considering data continuity and indicator representativeness, datasets covering the period from 2000 to 2024 were selected for subsequent analysis. For ecological indicators, calculations were conducted within the alpine grassland extent on the Qinghai-Tibet Plateau using spatially explicit remote-sensing and geospatial datasets. For socioeconomic indicators, prefecture-level statistics were used because prefecture-level administrative units are the basic governance and policy implementation units in the study region, and their socioeconomic conditions directly affect grassland management, ecological protection investment, infrastructure provision, and pastoral livelihood adaptation. Although this design may introduce some scale mismatch between ecological and socioeconomic variables, it reflects the actual administrative and governance context in which alpine grassland social-ecological processes occur. This limitation is acknowledged in the discussion.

3.2 Research methods

3.2.1 Catastrophe progression method

The catastrophe progression method originates from catastrophe theory and uses potential functions as the analytical basis to describe abrupt changes in complex systems. By incorporating interactions between state variables and control parameters, this method characterizes nonlinear system behaviors, where transition pathways among system states are governed by the membership functions of control variables. Integrating multi-valued fuzzy evaluation theory with nonlinear perspectives from complexity science, the catastrophe progression method is particularly suitable for analyzing systems characterized by the coexistence of gradual evolution and abrupt transitions, especially when internal mechanisms are difficult to explicitly determine. Common catastrophe models and their corresponding potential functions are presented in Table 2. Based on the established resilience assessment indicator system, the overall resilience objective was hierarchically decomposed according to the indicator structure. Appropriate catastrophe models were selected according to the number of control variables at each level. The standardized indicator values were then substituted into the corresponding potential functions, and membership values were calculated recursively through normalization procedures to quantify resilience of the alpine grassland social-ecological system on the Qinghai-Tibet Plateau (Xie et al., 2023). To address the problem of excessive concentration and insufficient discrimination of initial membership values, a discretization-based improvement was further introduced to enhance the differentiation among membership values and improve the robustness of resilience assessment results (Xu and Wang, 2024).

Table 2. Normalization equations for commonly used catastrophe models.

a, b, c, and d denote the control variables of the system; x_a, x_b, x_c, and x_d represent the corresponding catastrophe membership values derived from these control variables.

3.2.2 Three-dimensional kernel density estimation method

The three-dimensional kernel density estimation (3D-KDE) method is a non-parametric approach that relies on the intrinsic distribution characteristics of the data. By constructing continuous density functions, this method characterizes the temporal distribution dynamics of variables and reveals changes in distributional concentration and dispersion. The height and number of density peaks reflect the degree of concentration and multi-polarization characteristics, respectively, while the extension of distribution tails indicates the evolution of regional disparities (Zhao et al., 2024).

In this study, three-dimensional kernel density plots were generated using MATLAB to investigate the dynamic evolution of absolute differences in the resilience of the alpine grassland social-ecological system across the Qinghai-Tibet Plateau. The three axes of the 3D-KDE plots represent year, absolute difference in resilience, and the corresponding kernel density estimate, respectively. The absolute difference was calculated as the deviation of the resilience score of each prefecture-level administrative unit from the annual mean resilience score:

(1) Dit=|Rit−R¯t|

where.

D_it denotes the absolute difference in resilience for administrative unit i in year t, R_it denotes the resilience score of administrative unit i in year t, and R¯t denotes the mean resilience score of all administrative units in year t.

The kernel density estimator is expressed as follows:

(2) f^h(x)=1nh∑i=1nK(x−Xih)

where f^h(x) is the estimated density at evaluation point x, X_i is the i-th observation,

n is the number of observations, h is the bandwidth parameter, and K(·) is the kernel function. A Gaussian kernel function was used in this study:

(3) K(u)=12πexp(−u22)

The bandwidth parameter was determined automatically using the default bandwidth selection procedure in MATLAB. A larger bandwidth generally produces a smoother density curve, whereas an excessively large bandwidth may weaken the ability to capture local distributional features.

3.2.3 Markov chain

The Markov chain originates from stochastic process theory. By constructing a state transition probability matrix, it reflects the transition paths and evolutionary trends of a system among different states. The diagonal elements of the state transition probability matrix represent the self-maintaining inertia of resilience levels, while the off-diagonal elements indicate the probabilities of transitions or declines between adjacent levels. To determine whether regional resilience exhibits “low-level lock-in” or the potential for “cross-level leaps” during its evolution, the Markov chain method is introduced (Jiang et al., 2025). The specific calculation formula for the transition probability is as follows:

(4) pabt,t+d=p{Xt+d=b|Xt=a}=∑t=2000+d2024nabt,t+d∑t=20002024−dnat

where pabt,t+d represents the probability that the resilience index transfers from level a in year t to level b in year t+d; nabt,t+d denotes the number of samples undergoing the transition from level a in year t to level b in year t+d.

3.2.4 Self-organizing feature map network model

The self-organizing feature map (SOFM) is an unsupervised artificial neural network-based clustering method proposed by Kohonen. It identifies intrinsic structures within high-dimensional datasets through competitive learning and self-organization mechanisms, while preserving the topological relationships among input samples. Unlike supervised learning algorithms, SOFM does not require predefined categories and can automatically detect potential clustering patterns through neuron competition, lateral inhibition, and adaptive learning (Zhang R. et al., 2025). Therefore, it is particularly suitable for identifying nonlinear relationships and complex configurations within high-dimensional social-ecological systems.

In this study, the SOFM model was applied to identify resilience capacity configurations of alpine grasslands on the Qinghai-Tibet Plateau. Nine indicators representing resistance, recovery, and adaptation capacities across ecological, economic, and social subsystems were selected as input variables. The input vector of each prefecture-level unit was expressed as:

(5) Xi=(xi1,xi2,…,xim)

where X_i represents the input vector of the i-th sample, x_ij denotes the value of the j-th resilience capacity indicator, and m represents the number of input indicators. In this study, m=9.

To eliminate the influence of different measurement units among indicators, all input variables were normalized using the min–max normalization method. For positive indicators:

(6) Zij=Xij−XjminXjmax−Xjmin

For negative indicators:

(7) Zij=Xjmax−XijXjmax−Xjmin

where Z_ij represents the normalized value of indicator j for sample i, and Xjmax and Xjmin denote the maximum and minimum values of indicator j, respectively.

During the competitive learning process, the Euclidean distance between the input vector and neuron weight vector was calculated to determine the best matching unit (BMU):

(8) dic=∑j=1m(Zij−Wcj)2

where d_ic represents the distance between sample i and neuron c; W_cj denotes the connection weight between neuron c and input variable j. The neuron with the minimum distance was selected as the winning neuron:

(9) c∗=argmin1≤c≤M(dic)

where c^∗represents the best matching unit corresponding to sample i.

After determining the winning neuron, the weight vectors of the winning neuron and its neighboring neurons were updated iteratively according to the self-organizing learning rule:

(10) Wcj(t+1)=Wcj(t)+η(t)hc,c∗(t)·[Zij(t)−Wcj(t)]

where W_cj(t) represents the weight value at iteration t, η(t) is the learning rate, and hc,c^∗(t) denotes the neighborhood function.

The neighborhood function was defined using the Gaussian function:

(11) hc∗c(t)=exp(−‖rc−rc∗‖22σ2(t))

where r_c and r_c* represent the spatial locations of neurons c and the winning neuron c^∗, respectively, and σ(t) denotes the neighborhood radius.

The learning rate and neighborhood radius gradually decreased with increasing iteration numbers:

(12) η(t)=η0(1−tT)

(13) σ(t)=σ0(1−tT)

where η₀ and σ₀ represent the initial learning rate and neighborhood radius, respectively; T represents the maximum number of training iterations.

After network convergence, samples were assigned to different resilience capacity configurations according to their corresponding winning neurons:

(14) Ci=c∗

where C_i represents the resilience capacity pattern category of sample i.

The SOFM model was implemented using MATLAB R2020b. The network was trained for 5,000 iterations, and the number of output-layer neurons was tested within the range of 3–9 clusters. The optimal topology structure was determined according to the clustering performance and interpretability of resilience capacity configurations.

To test the robustness of the resilience assessment, we conducted sensitivity analyses using alternative normalization, weighting, and aggregation methods. Specifically, Z-score normalization was used as an alternative to min-max normalization, entropy weighting or equal weighting was used as an alternative weighting scheme, and linear weighted aggregation or geometric mean aggregation was used as an alternative aggregation method. The resulting resilience scores, classification results, and rankings of prefecture-level administrative units were compared across methods. The results showed that the main spatial patterns and resilience-level distributions remained generally stable, indicating that the assessment results were robust.

3.2.5 XGBoost-SHAP analysis

XGBoost is an ensemble learning algorithm that combines multiple decision trees through a gradient-boosting procedure. For the t-th boosting iteration, the model can be expressed as:

(15) y^i(t)=y^i(t−1)+ft(xi)

where y^i(t) is the predicted resilience value for observation i after t iterations, x_i is the vector of external drivers, and f_t represents the newly added regression tree. The model minimizes the following regularized objective function:

(16) ℒ=∑i=1nl(yi,y^i)+∑t=1TΩ(ft)

where 1(yi,y^i) is the loss function, T is the number of trees, and Ω(f_t) is the regularization term controlling model complexity. The regularization term helps reduce overfitting and improves the generalization performance of the model (Chen and Guestrin, 2016).

After model training, SHAP was used to interpret the contribution of each predictor to the model output. Based on cooperative game theory, the prediction for observation i can be decomposed as:

(17) f(xi)=ϕ0+∑j=1pϕij

where f(x_i) is the predicted comprehensive resilience, ϕ₀ is the mean model output, p is the number of predictors, and ϕ_ij is the SHAP value of predictor j for observation i. The SHAP value represents the marginal contribution of a variable to an individual prediction relative to the expected model output (Schmalhofer, 2019).

The global importance of each external driver was measured using its mean absolute SHAP value:

(18) Ij=1n∑i=1n|ϕij|

Where I_j denotes the global importance of predictor j. A larger I_j​indicates a greater contribution to the spatial and temporal variation in comprehensive resilience. The relative importance of each predictor was calculated as:

(19) RIj=Ij∑j=1pIj×100%

The sign of the SHAP value was used to determine the direction of the contribution. A positive SHAP value indicates that the variable increases the predicted resilience relative to the average prediction, whereas a negative SHAP value indicates a reduction in predicted resilience. SHAP dependence plots were further used to examine the nonlinear marginal effects and potential threshold ranges of key external drivers. In addition, SHAP interaction values were calculated to assess whether the contribution of one driver varied according to the level of another driver.

The model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE):

(20) R2=1−∑i=1n(yi−y^i)2∑i=1n(yi−y¯)2

(21) RMSE=1n∑i=1n(yi−y^i)2

(22) MAE=1n∑i=1n|yi−y^i|

Here, y_i, y^i, and y¯ denote the observed resilience value, predicted resilience value, and mean observed resilience value, respectively. It should be noted that SHAP values describe the model-based contribution and association of each external driver; they should not be interpreted as definitive causal effects. The equations used in the methodological analyses are presented in Equations 1–22.

4 Results

4.1 Spatiotemporal evolution of resilience of the alpine grassland social-ecological system patterns on the Qinghai-Tibet Plateau

The economic, ecological, social, and comprehensive resilience indices of alpine grassland social-ecological systems on the Qinghai-Tibet Plateau were evaluated for six benchmark years 2000, 2005, 2010, 2015, 2020, and 2024 to examine their spatiotemporal differentiation (Figure 2). Based on uniform thresholds applied across all years, resilience was classified into five levels: low, relatively low, medium, relatively high, and high. This consistent classification ensured comparability across years. Accordingly, the resilience-level changes presented in Figure 2 represent transitions between adjacent benchmark years only.

Regarding economic resilience, alpine grassland systems on the Qinghai-Tibet Plateau exhibited a stepwise improvement from 2000 to 2024, accompanied by pronounced north-south disparities. In 2000, economic resilience was generally low, with most areas characterized by “low” values, while only a few scattered areas in the eastern region reached the “relatively low” level. By 2005, areas with “low” economic resilience had substantially decreased, and most regions except Haixi Prefecture in Qinghai Province reached the “relatively low” level. In 2010, economic resilience further improved, with extensive areas developing toward the “medium” level, except for Haixi Prefecture and Yushu Prefecture in Qinghai Province, and the north-south differentiation pattern began to emerge.During 2015-2020, core urban areas and regions with concentrated industrial activities, including Xining City in Qinghai Province, Lasa City in Tibet, and Ganzi Prefecture in Sichuan Province, were the first to form contiguous areas characterized by “relatively high” to “high” economic resilience. In contrast, high-altitude pastoral regions, such as Haixi Prefecture in Qinghai Province and Naqu City in Tibet, exhibited relatively slower improvements. Overall, the spatial pattern gradually developed into a distribution characterized by “higher in the east and lower in the west, higher in the south and lower in the north.” By 2024, economic resilience was dominated by “relatively high” and “high” levels, with only Haibei Prefecture declining to the “medium” level. Although regional disparities gradually narrowed due to improvements in lagging areas, significant north-south differences remained (Figures 2b₁–b₆). The kernel density curves of economic resilience consistently exhibited a unimodal distribution, with the main peak continuously shifting toward higher values and the distribution range gradually expanding. In 2000, the main peak was concentrated in the low-value range, and the narrow distribution indicated a high degree of homogeneity among regions. Over time, the peak gradually shifted toward higher resilience levels, while peak height initially increased and subsequently declined. Meanwhile, the distribution range expanded and the right tail became increasingly extended, indicating the gradual emergence of disparities in economic resilience among different regions (Figures 3a₁, a₂).

Regarding ecological resilience, alpine grassland ecosystems on the Qinghai-Tibet Plateau exhibited a steady improvement from 2000 to 2024. In 2000, areas with “low” and “relatively low” ecological resilience dominated the region, indicating an overall weak ecological resilience level. During 2005-2010, Yushu Prefecture in Qinghai Province was the first area to reach the “high” resilience level, while regions surrounding Xining City, Rikaze City in Tibet, and Ganzi Prefecture in Sichuan Province improved to the “relatively high” level. In contrast, improvements in Ali Prefecture and Naqu City of Tibet were comparatively slower. During 2015-2020, areas initially characterized by “relatively high” and “high” ecological resilience gradually expanded outward, driving the formation of contiguous high-resilience regions. By 2024, the spatial pattern had stabilized, with “relatively high” and “high” resilience areas becoming dominant, while only Ali Prefecture in Tibet remained at a medium resilience level (Figures 2c₁–c₆). The kernel density curves of ecological resilience evolved from a bimodal distribution toward a unimodal distribution, with the main peak continuously shifting to the right. In the early stage, the distribution exhibited a distinct secondary peak in the high-value range, indicating the presence of localized high-resilience areas. As the evolution proceeded, the main peak gradually shifted toward higher resilience levels, while the secondary peak progressively merged into the main peak, resulting in a unimodal distribution pattern. Meanwhile, the distribution range became narrower, suggesting that initially high-resilience regions exerted a diffusion effect on surrounding areas and promoted overall improvements in alpine grassland ecosystem stability across the Qinghai-Tibet Plateau (Figures 3b₁, b₂).

Regarding social resilience, alpine grassland regions on the Qinghai-Tibet Plateau exhibited a continuous and balanced improvement from 2000 to 2024, reflecting increasingly inclusive development in social livelihood conditions. In 2000, the entire region was characterized by a “low” resilience level, with only a few areas reaching the “relatively low” level. During 2005-2015, all regions generally improved to the “relatively high” level, while areas including Xining City in Qinghai Province, Lasa City, Rikaze City, and Changdu City in Tibet, as well as Ganzi Prefecture in Sichuan Province, were the first to achieve the “high” resilience level. By 2020, the entire region had stabilized at the “relatively high” to “high” levels. In 2024, all regions except Guoluo Prefecture in Qinghai Province reached the “high” resilience level (Figures 2d₁–d₆). The kernel density curves of social resilience exhibited a narrow-peaked distribution pattern with a continuous rightward shift. The curves maintained a steep unimodal shape with a narrow distribution range, while peak heights continuously increased, indicating that social resilience levels across the Qinghai-Tibet Plateau became increasingly concentrated with reduced spatial dispersion. The main peak shifted steadily toward higher values along the horizontal axis, and the curves showed almost no low-value tail, suggesting that social resilience experienced a stable and balanced improvement across the entire region (Figures 3c₁-c₂).

Regarding comprehensive resilience, alpine grassland systems on the Qinghai-Tibet Plateau exhibited a continuous and steady improvement from 2000 to 2024, indicating an overall enhancement in regional resilience. In 2000, areas with “low” comprehensive resilience dominated the plateau, while Xining City and Hainan Prefecture in Qinghai Province showed relatively higher resilience levels and gradually expanded toward the southern regions. During 2005-2010, comprehensive resilience gradually increased to the “medium” level, with Xining City and Hainan Prefecture in Qinghai Province, Lasa City in Tibet, and Ganzi Prefecture in Sichuan Province being the first areas to reach the “relatively high” level. During 2015-2020, regions with “relatively high” and “high” comprehensive resilience expanded continuously and formed contiguous high-resilience areas. By 2024, “high” resilience areas became dominant across the Qinghai-Tibet Plateau, with only Ali Prefecture in Tibet and Haibei Prefecture in Qinghai Province remaining at the “relatively high” level (Figures 2a₁–a₆). The kernel density curve of comprehensive resilience exhibited a continuous rightward shift, with the peak height generally increasing despite some fluctuations during intermediate periods. The distribution maintained a unimodal pattern, indicating that comprehensive resilience of alpine grassland systems on the Qinghai-Tibet Plateau experienced a sustained and stable upward trajectory (Figure 3d₁–d₂).

4.2 Dynamic Transition Patterns of resilience of the alpine grassland social-ecological system on the Qinghai-Tibet Plateau

To reveal the dynamic transition processes and evolutionary patterns of comprehensive, economic, ecological, and social resilience of alpine grasslands on the Qinghai-Tibet Plateau from 2000 to 2024, the resilience levels were discretized into five categories: Level I, Level II, Level III, Level IV, and Level V, representing low, relatively low, medium, relatively high, and high resilience levels, respectively. A Markov transition probability matrix was constructed to characterize the dynamic transitions among different resilience states. The main diagonal elements of the matrix represent the probabilities of maintaining the original resilience levels, whereas the off-diagonal elements indicate the probabilities of upward transitions or downward regressions between different resilience levels.

From 2000 to 2024, resilience of the alpine grassland social-ecological system on the Qinghai-Tibet Plateau exhibited significant club convergence characteristics. The transition probabilities along the diagonal were consistently higher than those of off-diagonal elements, while non-zero transition probabilities were mainly concentrated around the adjacent positions of the diagonal. The probabilities of transitions across two or more resilience levels approached zero, indicating that resilience evolution primarily occurred through gradual transitions between neighboring levels, without substantial leapfrog improvements or abrupt declines (Table 3). Based on the average self-maintenance probabilities of diagonal elements, the stability of different resilience dimensions followed the order of social resilience > economic resilience > ecological resilience, with all values exceeding 78.848%. The average self-maintenance probability of comprehensive resilience was 82.0%, suggesting a high likelihood of maintaining the original resilience state over time. At different resilience levels, the high-resilience level (Level V) exhibited the strongest locking effect, with self-maintenance probabilities ranging from 93.3% to 96.5%. This indicates that once a region entered the high-resilience state, it was highly likely to remain at that level in the following period, with only a limited probability of declining to lower levels. The pronounced convergence of high-resilience clubs was particularly evident. Among the three subsystems, social resilience showed the highest self-maintenance probability at the high-resilience level, indicating the strongest stability. The low-resilience level (Level I) also demonstrated strong persistence, with self-maintenance probabilities exceeding 77%. Only 17.9%-22.9% of regions transitioned to the relatively low-resilience level (Level II), while upward transitions to medium or higher resilience levels remained difficult, suggesting the potential existence of a low-resilience equilibrium trap. Among the subsystems, ecological resilience exhibited the highest self-maintenance probability at the low-resilience level, indicating stronger persistence of ecological vulnerability. The relatively low-resilience level (Level II) showed substantial differentiation among subsystems. The self-maintenance probabilities of Level II for comprehensive and social resilience were both approximately 80%, compared with 75.5% for economic resilience and only 69.9% for ecological resilience, indicating that ecological resilience at this stage was more unstable and more likely to undergo state transitions. The medium-resilience level (Level III) had the lowest overall self-maintenance probability and was relatively unstable, with approximately 21.0%-31.4% probability of upward or downward transitions. Among them, Level III ecological resilience showed the weakest stability. The relatively high-resilience level (Level IV) exhibited relatively balanced stability, with self-maintenance probabilities ranging from 77.7% to 79.0%. Regarding positive resilience transitions, improvements generally occurred through stepwise progression between adjacent levels. Economic resilience at Level I exhibited the greatest potential for upward transition, whereas Level I and Level IV ecological resilience showed the strongest barriers to upward improvement. During the transition from Levels II to III, ecological resilience demonstrated the highest probability of positive advancement. Regarding downward transition risks, the overall probability of resilience decline remained relatively low. However, ecological resilience exhibited higher degradation risks than economic, comprehensive, and social resilience across most levels. In particular, the probability of downward transition for Level III ecological resilience reached 7.6%, while the downward transition probability of Level IV ecological resilience was three to six times higher than that of other resilience dimensions. Economic resilience showed a small but persistent possibility of downward regression across all resilience levels.

Table 3. Markov transition probability matrix of resilience of the alpine grassland social-ecological system levels on the Qinghai-Tibet Plateau from 2000 to 2024.

Levels I, II, III, IV, and V represent low, relatively low, medium, relatively high, and high resilience levels, respectively. t/(t+1) denotes the transition probability from year t to year t+1.

As shown in Figure 4, resilience-level transitions were compared across adjacent benchmark years, namely 2000–2005, 2005–2010, 2010–2015, 2015–2020, and 2020–2024, and were classified into three types based on actual changes in the resilience index: upward, stable, and downward. Among the prefecture-level administrative units on the Qinghai-Tibet Plateau, 51% exhibited an overall staged upward trend from 2000 to 2024, 32% showed a downward trend, and 17% remained stable. Across the four resilience dimensions, upward trajectories were dominant. Specifically, 40% of the administrative units showed an increase in comprehensive resilience, compared with 32% showing a decrease and 28% remaining stable. The proportions of administrative units with increasing economic, ecological, and social resilience were 51%, 54%, and 47%, respectively, while the corresponding proportions with declining trajectories were 36%, 35%, and 32%. Stable trajectories accounted for 13% of economic resilience, 11% of ecological resilience, and 21% of social resilience. Overall, ecological resilience had the highest proportion of upward trajectories, whereas comprehensive resilience had the highest proportion of stable trajectories. These results are generally consistent with the overall improvement in resilience reported in Section 4.1, although some administrative units still exhibited declining or stable trajectories during the study period.

4.3 Evolutionary characteristics of resilience of the alpine grassland social-ecological system on the Qinghai-Tibet Plateau

Previous analyses demonstrated that resilience of the alpine grassland social-ecological system on the Qinghai-Tibet Plateau exhibited significant spatiotemporal heterogeneity. Alpine grassland social-ecological resilience is an emergent property resulting from the dynamic coupling of economic, social, and ecological subsystem elements over time, with complex interactions formed through internal self-organization processes. Due to inherent differences in resource endowments, industrial foundations, and external disturbance conditions among prefecture-level administrative units, differentiated resilience capacity configurations have emerged, which subsequently influence regional overall resilience levels and long-term evolutionary trajectories. Therefore, clarifying the evolutionary patterns of regional resilience capacities and coordinating interactions among different subsystems are essential for enhancing comprehensive resilience on the Qinghai-Tibet Plateau.

4.3.1 Resilience capacity configuration patterns on the Qinghai-Tibet Plateau

In this study, the SOFM clustering algorithm was employed to identify the combinations of Alpine grassland social-ecological resilience capacity configurations. The optimal number of clusters was determined to be five, and each cluster was named according to its dominant resilience capacity characteristics: adaptation-dominated type (Type I), recovery-dominated type (Type II), economically recovery-constrained type (Type III), ecological resistance–economic resistance dominated type (Type IV), and social resistance–economic recovery dominated type (Type V). Figure 5 illustrates the temporal evolution of the proportions of the five resilience capacity configuration patterns across prefecture-level administrative units on the Qinghai-Tibet Plateau from 2000 to 2024. Type V accounted for the largest proportion, with an average share of 28.76%, and remained dominant throughout the study period. Its proportion showed an overall increasing trend with fluctuations, reaching a maximum of approximately 43% in 2024. Type I ranked second, with an average proportion of 26.29%, characterized by an initial rapid increase followed by long-term fluctuations at a relatively high level, without a persistent downward trend. The average proportions of Type II and Type III were 15.24% and 15.81%, respectively. Type IV had the lowest average proportion (13.90%), showing the poorest stability and the strongest fluctuations among all categories. Its proportion declined to zero in 2020 and remained at a low level of only 4.76% in 2024. Overall, resilience evolution on the Qinghai-Tibet Plateau was primarily dominated by two configuration patterns: social resistance–economic recovery dominated resilience and adaptation-oriented resilience, indicating that the long-term improvement of regional resilience largely depended on the coordinated development of social-economic capacities and adaptive capabilities.

For specific prefecture-level administrative units, Lasa City and Changdu City in Tibet consistently remained in Class I, while Xining City and Hainan Prefecture in Qinghai Province, as well as Kashi Prefecture in Xinjiang, maintained a stable Class I status throughout the study period. Shannan City in Tibet remained in Class II, and Huangnan Prefecture in Qinghai Province remained in Class III. Guoluo Prefecture in Qinghai Province consistently belonged to Class V, while Haixi Prefecture in Qinghai Province, Ali Prefecture and Naqu City in Tibet, and Kezhou Autonomous Prefecture in Xinjiang maintained a stable Class V status. From 2000 to 2024, Aba Prefecture in Sichuan Province experienced transitions between Class II and Class III, with Class II being the dominant pattern. Diqing Prefecture in Yunnan Province, Ganzi Prefecture in Sichuan Province, Haibei Prefecture in Qinghai Province, Linzhi City and Rikaze City in Tibet were primarily characterized by Class III resilience patterns. Gannan Prefecture in Gansu Province was mainly associated with Class IV, whereas Haidong City and Yushu Prefecture in Qinghai Province were predominantly classified as Class V, while also exhibiting transitions toward other resilience classes.

From the perspective of spatial patterns, the five resilience capacity configurations exhibited significant geographical clustering characteristics. The adaptation-dominated pattern (Class I) was highly concentrated in the northeastern and southwestern marginal regions of the Qinghai-Tibet Plateau, including Xining City and Hainan Prefecture in Qinghai Province, as well as Changdu and Lasa cities in Tibet. These areas generally possess relatively higher levels of urbanization, stronger policy responsiveness, and greater institutional innovation capacity. The socioeconomic synergy pattern (Class V) was widely distributed across the western alpine pastoral regions, including Haixi and Guoluo prefectures in Qinghai Province, Linzhi and Naqu cities in Tibet, and Kezhou Autonomous Prefecture in Xinjiang. This pattern exhibited a broad contiguous spatial distribution, with resilience enhancement largely relying on the coordinated advancement of ecological compensation mechanisms and livestock husbandry transformation. The economic recovery lagging pattern (Class III) was mainly scattered across the transitional zone between western Sichuan and northern Yunnan, including Ganzi and Aba prefectures in Sichuan Province and Diqing Prefecture in Yunnan Province. In these regions, relatively weak industrial foundations and pronounced geographical constraints limited economic recovery capacity. Overall, the spatial differentiation of resilience capacity configurations showed a strong consistency with the Hu Huanyong Line and altitudinal gradients. The long-term constraints imposed by geographical backgrounds have shaped differentiated resilience capacity configurations through variations in resource endowments and market accessibility.

4.3.2 Relationships among resilience capacities and their effects on resilience levels

The formation and evolution of resilience capacity patterns in prefecture-level administrative units across the Qinghai-Tibet Plateau can ultimately be traced back to the integration and allocation of underlying elements, which are manifested through the interactions among specific resilience capacities. In this study, Spearman correlation coefficients were employed to reveal the synergistic and trade-off relationships among resilience capacities within ecological, economic, and social subsystems (Figure 6). Combined with the previously measured resilience levels, the differentiated evolutionary mechanisms of Alpine grassland social-ecological resilience were further explored across different city types. From the perspective of the overall interaction structure, all city types exhibited strong resilience coupling characteristics. Specifically, economic recovery capacity, economic adaptation capacity, and social adaptation capacity, as well as ecological recovery capacity and ecological adaptation capacity, demonstrated highly significant positive correlations. These relationships indicate that, when responding to external disturbances, recovery and adaptation processes exhibit strong systemic linkages and coordinated development characteristics. However, the social resistance capacity generally showed negative correlations with both ecological resistance capacity and economic resistance capacity, suggesting potential resource competition or structural conflicts among different subsystems during the resistance stage. Further analysis of the differentiated characteristics among various city types revealed that Class I prefecture-level administrative units exhibited the densest positive interactions among resilience dimensions and the highest degree of resilience coordination, indicating a highly integrated resilience system. In contrast, Class II cities showed the fewest significant correlations within the interaction matrix, suggesting relatively weak linkages among subsystems and an underdeveloped coupling mechanism. Class III prefecture-level administrative units demonstrated prominent negative structural conflicts, particularly the strong negative correlations between social resistance capacity and both ecological resistance capacity and economic resistance capacity, indicating constraints and imbalances in fundamental defensive capabilities. Meanwhile, Class IV and Class V prefecture-level administrative units were generally located in an intermediate transitional stage. Among them, Class V prefecture-level administrative units also exhibited relatively pronounced conflicts between social resistance capacity and ecological resistance capacity. Overall, prefecture-level administrative units at different resilience levels exhibited distinct operational mechanisms during resilience evolution. Therefore, uniform governance strategies are unlikely to achieve optimal outcomes. Instead, future resilience enhancement efforts should adopt differentiated regulation pathways tailored to local conditions, focusing respectively on breaking subsystem isolation, alleviating structural conflicts, and maintaining synergistic interactions among resilience capacities. Figure 6 presents Spearman rank correlation coefficients calculated from prefecture-year observations over the study period and is intended to describe the overall association patterns among different resilience-capacity dimensions. Because the same prefecture-level administrative units are repeatedly observed across years, and because common temporal trends and spatial dependence may be present, the correlation coefficients and their statistical significance should not be regarded as strict inferential results under an independent and identically distributed assumption. Therefore, the results in Figure 6 are presented only as descriptive evidence and are not used to support causal or mechanistic conclusions.

To further identify the external driving mechanisms of comprehensive resilience, an XGBoost model combined with SHAP analysis was applied. In this analysis, the comprehensive resilience index was used as the dependent variable, while annual mean precipitation, annual mean temperature, extreme precipitation frequency, mean elevation, slope, per capita GDP, urbanization rate, and nature reserve coverage were used as predictors. These variables were selected because they represent key exogenous climatic, topographic, socioeconomic, and policy-related influences, and they do not overlap with the internal dimensions used to construct the resilience index.

The XGBoost model showed satisfactory predictive performance, with an R² of 0.768, RMSE of 0.058, and MAE of 0.044. According to the mean absolute SHAP values, annual mean precipitation was the most important external driver of comprehensive resilience, followed by mean elevation, annual mean temperature, and nature reserve coverage (Table 4, Figure 7). This ranking suggests that hydrothermal conditions and topographic constraints play a dominant role in shaping regional resilience patterns. Nature reserve coverage also showed a relatively high contribution, indicating that ecological protection policy exerts a substantial positive influence on comprehensive resilience.

Table 4. SHAP-based attribution of external drivers to comprehensive resilience.

The dependent variable is the comprehensive resilience index. The predictors include only exogenous climatic, topographic, human-activity-related, and policy-related variables that were not used in constructing the resilience index. The nine resilience capacity dimensions were excluded from the predictor set to avoid circularity. Mean |SHAP| values indicate the average absolute contribution of each variable to the model output. Relative importance was calculated as the proportion of each variable’s mean |SHAP| value in the total mean |SHAP| values of all predictors. Nonlinear and threshold patterns were identified based on SHAP summary plots and SHAP dependence plots. Model performance: R² = 0.768, RMSE = 0.058, MAE = 0.044. Sample size: N = 525.

In contrast, slope and extreme precipitation frequency exhibited negative effects, implying that steeper terrain and more frequent climatic extremes may weaken the stability and recovery potential of the regional system. Human-activity-related variables, including urbanization rate and per capita GDP, showed moderate importance and mixed effects, suggesting that socioeconomic development may enhance resilience at moderate levels but may also generate ecological pressure when development intensity becomes excessive. The SHAP dependence plots further revealed clear nonlinear response patterns (Figure 8). For example, the positive contribution of precipitation became more pronounced after a moderate threshold was reached, while the effect of temperature shifted from beneficial to inhibitory under higher warming conditions. Similarly, nature reserve coverage displayed a stronger positive effect once ecological protection reached a certain intensity (Figure 9). Overall, these findings indicate that the drivers of comprehensive resilience operate through nonlinear and interactive mechanisms rather than purely linear relationships.

5 Discussion

5.1 Spatiotemporal differentiation and transition patterns of resilience evolution

The comprehensive resilience of alpine grasslands on the Qinghai-Tibet Plateau exhibited a steady increasing trend; however, the evolutionary trajectories of the three major subsystems showed pronounced differentiation. Economic resilience demonstrated a stepwise improvement accompanied by significant north-south disparities, reflecting the uneven spatial distribution of market accessibility, industrial foundations, and policy support (Cao et al., 2018; Li et al., 2026). Ecological resilience evolved from a bimodal to a unimodal distribution, suggesting that the radiation and the distribution showed a tendency toward convergence of localized high-resilience cores (e.g., Yushu Prefecture) have gradually emerged, allowing the positive externalities of ecological governance to diffuse spatially. Nevertheless, peripheral regions such as Ngari Prefecture remained locked in a medium-resilience state. Social resilience exhibited a balanced narrow-peaked improvement pattern, highlighting the characteristics of inclusive development, which are closely associated with the continued promotion of equalization policies in public services, including education, healthcare, and social security (Guo et al., 2024; Li et al., 2025). This asynchronous evolution supports the theoretical perspective that resilience does not increase synchronously across systems, and further indicates that under the combined influences of climate change and policy interventions, ecological recovery tends to exhibit time lags and the distribution showed a tendency toward convergence, whereas social systems generally respond more synchronously due to stronger institutional dependence (Feng et al., 2021; Tatham and Lemaire, 2025; Brown, 2026).

The Markov chain analysis further revealed significant club convergence characteristics in resilience evolution. The self-maintenance probability of high-resilience states exceeded 93%, while low-resilience states exhibited substantial barriers to transitioning toward medium or higher levels. Transitions were almost entirely restricted to adjacent resilience levels, with cross-level leaps rarely occurring. This pattern can be interpreted through the perspectives of path dependence and cumulative capacity development (Liu, 2025). Regions with high resilience generally possess stronger economic foundations, favorable ecological conditions, and well-established social service networks, thereby forming self-reinforcing feedback loops. In contrast, low-resilience regions are constrained by structural limitations, including harsh natural environments, limited industrial diversity, and weak fiscal capacity. Even with external policy support, these regions may experience considerable difficulty achieving substantial improvements in resilience levels within a short period. Notably, ecological resilience exhibited the highest persistence probability at low resilience levels compared with economic and social resilience, indicating that ecosystem recovery processes are relatively slow and highly sensitive to disturbances. Once degraded into a low-resilience state, ecological systems may require substantially longer recovery periods than economic or social systems (Jones and Schmitz, 2009; Biswas and Biswas, 2026). However, the extremely high persistence probability of high-resilience states (>93.3%) may not solely reflect genuine path dependence. The possibility of a ceiling effect within the evaluation framework should also be considered. Specifically, if the upper limits of resilience indicators are insufficiently flexible, high-resilience regions experiencing minor degradation may show limited declines in measured resilience scores. This issue may be particularly relevant for social resilience indicators (e.g., years of schooling and social security coverage), where many regions approached saturation after 2020, reducing their discriminatory power. Therefore, future studies should incorporate high-frequency ecological sensitivity indicators derived from satellite observations, such as vegetation drought resistance and snowline fluctuations, to overcome upper-bound constraints of current indicators and more accurately characterize emerging risks faced by high-resilience regions.

5.2 Structural differentiation and synergy-trade-off mechanisms of resilience capacity configurations

The five resilience capacity patterns identified through SOFM clustering revealed the underlying structural mechanisms driving differentiated Alpine grassland social-ecological resilience trajectories. The adaptation-dominated pattern (Class I) and the socioeconomic synergy pattern (Class V) remained dominant over the study period, indicating that adaptive capacity in response to future uncertainties and cross-system coordination are key drivers of resilience enhancement on the Qinghai-Tibet Plateau. However, the emergence of the economic recovery lagging pattern (Class III) and the ecological resistance-economic resistance dominated pattern (Class IV) highlights structural mismatches between fundamental defensive capacity and recovery capacity in some regions. The distinct temporal evolution and spatial distribution characteristics of different resilience patterns reflect how variations in resource endowments, industrial foundations, and external disturbance regimes among prefecture-level administrative units have been transformed into differentiated resilience configurations through self-organizing processes within social-ecological systems. Correlation analysis further demonstrated strong synergistic relationships between recovery and adaptation capacities within both ecological and economic subsystems, whereas social resistance capacity generally showed negative associations with ecological and economic resistance capacities. This structural conflict during the resistance stage may result from competition among subsystems under conditions of limited resources. For example, maintaining certain grazing intensities to ensure basic livelihoods in pastoral areas may weaken ecosystem resistance capacity (Shen et al., 2025); similarly, the expansion of mining and tourism industries to sustain local fiscal revenues may generate direct disturbances to alpine grassland ecosystems. Further examination of different resilience capacity patterns revealed that Class I prefecture-level administrative units exhibited the strongest positive interactions among resilience dimensions and the highest level of coordination, indicating a relatively integrated resilience structure. In contrast, Class II prefecture-level administrative units showed immature coupling mechanisms among subsystems, suggesting insufficient interactions and coordination. Class III prefecture-level administrative units experienced significant constraints and imbalances in fundamental defensive capacities, whereas Class IV and Class V prefecture-level administrative units remained in an intermediate transitional stage, with particularly pronounced conflicts between social resistance and ecological resistance capacities (Hoekstra et al., 2018; Zhang et al., 2024).These synergy-trade-off relationships suggest that resilience governance should not pursue the optimization of a single subsystem in isolation. Instead, effective resilience enhancement requires identifying and managing key conflicts among subsystems while promoting coordinated improvements across ecological, economic, and social dimensions.

5.3 Differentiated regulation pathways and methodological reflections

Based on the above findings, this study suggests that enhancing resilience of the alpine grassland social-ecological system on the Qinghai–Tibet Plateau should follow the principles of “classified strategies, gradual breakthroughs, and coordinated governance.” Differentiated regulation strategies should be developed according to specific resilience capacity patterns. prefecture-level administrative units characterized by the adaptation-dominated pattern should consolidate their advantages in self-adaptive capacity by strengthening institutional innovation and social capital development. prefecture-level administrative units with a recovery-dominated pattern should improve post-disturbance response efficiency and enhance technical support for ecological restoration. prefecture-level administrative units exhibiting the economic recovery lagging pattern should prioritize the improvement of industrial foundations and infrastructure deficiencies. Meanwhile, cities characterized by the ecological-economic resistance-dominated pattern should avoid the imbalance of “excessive defense but insufficient recovery” by increasing investments in restoration-oriented capacities (Dong and Wang, 2026). Furthermore, to address the problem of low-resilience state lock-in, intervention strategies should adopt a stepwise transition pathway rather than pursuing rapid leapfrog improvements. Through phased goal setting, dynamic monitoring, and adaptive adjustments, low-resilience prefecture-level administrative units can gradually escape from the low-level equilibrium trap and achieve progressive resilience enhancement (Storbjörk and Hjerpe, 2024). In addition, greater attention should be paid to the management of synergies and trade-offs among subsystems. Through institutional arrangements (e.g., integrating ecological compensation programs with livelihood transition training for pastoral communities), spatial planning (e.g., delineating strictly protected zones and sustainable utilization areas), and technological innovation (e.g., smart grazing systems and grassland health monitoring), structural conflicts among subsystems can be transformed into coordinated development pathways.

From a methodological perspective, this study integrated catastrophe progression analysis, three-dimensional kernel density estimation, Markov chain analysis, SOFM clustering, and XGBoost-SHAP to characterize the nonlinear, multidimensional, and dynamically interactive features of resilience in alpine grassland social-ecological systems. This integrated framework provides a methodological reference for resilience assessment in similar regions. In particular, the XGBoost-SHAP results extend recent ecosystem-health studies that distinguish climate change from human activities and examine nonlinear interactions. For example, Cheng et al.(2025) demonstrated that climate and human drivers jointly shape cropland ecosystem health through heterogeneous and nonlinear mechanisms. Consistent with this line of research, our findings show that comprehensive resilience is not driven by a single factor but is jointly shaped by climatic, topographic, socioeconomic, and policy-related external drivers through threshold and interaction effects (Cheng et al., 2025). Nevertheless, several limitations remain. First, the conceptual boundaries of several indicators require further refinement. For example, the growth rate of fiscal expenditure on ecological protection represents an external policy intervention rather than an intrinsic ecological recovery process. Similarly, ecological compensation coverage, social security coverage, and policy awareness primarily reflect enabling conditions or governance support for resilience. Although these indicators are relevant to the functioning of the alpine grassland social-ecological system, they should be interpreted as proxies for resilience-related capacity rather than direct measurements of realized recovery or adaptation. Future studies should distinguish more clearly between endogenous system capacities, external interventions, and observed post-disturbance responses. Second, limitations exist in data scale and temporal resolution. Due to data availability constraints, some indicators were aggregated at the prefecture-level city scale, making it difficult to capture micro-scale resilience differences among counties or pastoral households. Moreover, resilience assessment was primarily based on annual datasets, which may not adequately capture short-term disturbances caused by extreme climate events and subsequent recovery processes. Third, although this study identified nonlinear associations between external drivers and comprehensive resilience using XGBoost-SHAP, the analysis does not establish causal relationships. Future studies could combine high-frequency remote-sensing observations, quasi-experimental designs, structural equation modeling, or system dynamics approaches to further examine the causal mechanisms underlying resilience evolution. Additionally, theoretical frameworks should distinguish between passive adaptation and proactive transformation, clarifying their respective roles in shaping long-term resilience trajectories.

6 Conclusions

Based on the “Components–Attributes–Capacities” analytical framework, this study integrated the catastrophe progression method, three-dimensional kernel density estimation, Markov chain analysis, SOFM clustering, and XGBoost-SHAP analysis to examine the spatiotemporal evolution, transition patterns, capacity configurations, and external drivers of prefecture-level alpine grassland social-ecological resilience on the Qinghai-Tibet Plateau. Using 21 prefecture-level administrative units on the Qinghai-Tibet Plateau from 2000 to 2024 as the study sample, the main conclusions are summarized as follows:

• The overall resilience of alpine grasslands on the Qinghai-Tibet Plateau exhibited a steady increasing trend; however, the evolutionary trajectories of the economic, ecological, and social subsystems showed significant differentiation. Economic resilience demonstrated a stepwise increase with pronounced north-south disparities, accompanied by a unimodal distribution shifting toward higher values. Ecological resilience evolved from a bimodal to a unimodal distribution, with increasing radiation effects from high-resilience regions, although marginal areas remained constrained by medium-resilience lock-in. Social resilience exhibited balanced and inclusive improvement, with a relatively stable overall development trajectory.
• resilience of the alpine grassland social-ecological system levels exhibited significant club convergence characteristics and path dependence. The self-maintenance probabilities of different resilience levels exceeded 78%, and high-resilience states showed the strongest locking effect, with persistence probabilities ranging from 93.3% to 96.5%. Low-resilience regions faced substantial barriers to improvement and mainly experienced transitions between adjacent levels, with limited potential for leapfrog advancement. Moreover, degraded low-resilience ecological systems exhibited substantially greater restoration difficulties than economic and social subsystems.
• Five resilience capacity configuration patterns were identified. The socioeconomic synergy pattern and the adaptation-dominated pattern were the dominant configurations, together accounting for more than 50% of the study area. However, some regions exhibited structural mismatches between fundamental capacity and recovery capacity. Meanwhile, resilience capacities showed both synergistic and trade-off relationships: recovery and adaptation capacities were highly coordinated, whereas social resistance capacity showed significant resource competition and structural conflicts with ecological and economic resistance capacities, with substantial differences among Alpine grassland social-ecological resilience types.
• Based on these findings, this study proposes that resilience enhancement on the Qinghai-Tibet Plateau should follow the principles of “classified strategies, gradual improvement, and coordinated governance.” Differentiated regulation approaches should be implemented according to regional resilience capacity patterns, with emphasis on the coordinated cultivation of recovery and adaptation capacities and the balancing of trade-offs among ecological, economic, and social subsystems. These findings provide theoretical insights and empirical evidence for maintaining ecological security and promoting sustainable development of pastoral regions on the Qinghai-Tibet Plateau.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Amira Sharief, Zoological Survey of India, India

Reviewed by: Peiheng Yu, Hong Kong Polytechnic University, Hong Kong SAR, China

Jian Chen, University College London, United Kingdom