Evolution of Ecological Vulnerability and Scenario Simulations in the Yellow River Source Region Under Climate Change

Abstract

Amid accelerating global environmental change, assessing ecological vulnerability is critical for sustainability science. Focusing on the Yellow River Source Region (YRSR)—a key water source and ecological shield in China—this study develops an integrated assessment system based on the “Pressure–State–Response” (PSR) framework, incorporating 29 indicators. A combined weighting approach integrating analytic hierarchy process (AHP) with entropy-based objective weighting characterizes the spatiotemporal patterns, drivers, and future trajectories of ecological vulnerability. Key findings reveal: (1) heterogeneous warming–wetting trends with stronger humidification in the south and relative stability in the north drive divergent hydrological responses, highlighting the limitations of single-climate metrics in explaining vulnerability dynamics; (2) vulnerability patterns are primarily shaped by climatic factors—especially temperature and potential evapotranspiration—with anthropogenic pressures serving as secondary modulators, reinforcing the foundational role of thermal and moisture regimes in alpine ecosystem resilience; and (3) scenario projections consistently identify the northeast as a persistently high-vulnerability zone, yet show that balanced socioeconomic development can reconcile ecological protection with development needs. Based on these insights, a four-tier ecological zoning scheme and a governance framework comprising three strategies—strict conservation, adaptive regulation, and sustainable utilization—are proposed. This work offers actionable scientific guidance for tailored ecological conservation in the YRSR and contributes methodological advancements for vulnerability assessment and adaptive management of high-elevation ecosystems globally.

1. Introduction

Amidst intensifying global environmental shifts [1,2], ecological deterioration has become a pressing transboundary issue [3,4]. Given the profound interdependence between socioeconomic systems and natural ecosystems, the assessment of ecological vulnerability has emerged as a central theme in sustainability science and climate adaptation research [5]. Ecological vulnerability reflects an intrinsic property of ecosystems—namely, their tendency to undergo state transitions under external pressure, as well as their capacity to absorb disturbances or recover to a reference state [6]. This assessment is directly linked to several core objectives of the United Nations 2030 Agenda for Sustainable Development, in particular SDG 13 (Climate Action), SDG 15 (Life on Land), and SDG 6 (Clean Water and Sanitation).

Situated at the geographic core of Qinghai Province, the YRSR serves as both a crucial freshwater reservoir and an ecological barrier for the entire Yellow River Basin. Its environmental health is vital to regional and national ecological security, directly supporting SDG 6 targets related to clean drinking water supply and sustainable water resource management at the basin scale [7]. However, the region is constrained by extreme abiotic factors: high altitude, persistent cold, hypoxia, rugged terrain, limited thermal energy, scarce moisture, low vegetation cover, simplified biotic communities, slow biogeochemical cycling, extremely low buffering capacity and self-restoration potential, as well as widespread soil erosion [8]. These natural limitations are further exacerbated by mounting anthropogenic pressures—including overgrazing, infrastructure expansion, and resource extraction—heightening its susceptibility to irreversible ecological decline [6]. Moreover, the YRSR harbors exceptional biodiversity and functions as a key regulator of hydro-climatic stability across western China and downstream areas. It has become a globally recognized site for research on climate sensitivity, ecosystem tipping points, and adaptive governance in high-altitude environments [9,10]. Its protection and sustainable development thus directly align with SDG 15 and SDG 13. In response, the Outline for Ecological Protection and High-Quality Development of the Yellow River Basin [11] explicitly mandates systematic monitoring and assessment of degradation trends and pollution dynamics across the basin.

Ecological vulnerability assessment aims to diagnose the bottlenecks that constrain regional sustainable development and provide evidence support for the design of protection interventions and restoration plans [12]. Existing research has made significant progress in methods, and widely applied theoretical frameworks include the PSR model, the SRP method, and the VRD framework [13,14,15]. Although research methods have been continuously enriched, there are still three key scientific gaps in the current ecological vulnerability assessment of the Yellow River Source Region. First, most studies rely on a few climate indicators, such as annual average temperature and annual precipitation, to explain the changes in vulnerability, ignoring the role of comprehensive hydrothermal factors such as potential evapotranspiration and water balance, and failing to reveal the spatial differentiation mechanism of vulnerability under the “warming and humidification” background—the understanding of this mechanism is crucial for achieving climate adaptation actions under the Sustainable Development Goal 13 [16]. Second, the relative contributions of climate factors and human pressure to vulnerability formation are often compared in parallel, but whether there is a hierarchical relationship of “climate lays the foundation pattern, and human imposes local regulation” between the two has not been quantitatively decomposed, and this directly relates to how to balance ecological protection and livelihood development under the framework of Goal 15. Third, future scenario predictions mostly focus on average state changes, and there is insufficient attention to the identification and management implications of “persistent high-vulnerability areas” under different social and economic development paths, which affects the regional resilience planning oriented towards sustainable cities and communities. In addition, existing research generally has problems such as scattered indicator design, poor spatial resolution, and failure to comprehensively consider natural and human factors and governance effects, which affect the accuracy of diagnosis.

Based on the above background, this study focuses on the YRSR, uses the PSR model, and through the refinement of the indicator system and the optimization of the spatial scale, constructs a comprehensive evaluation system consisting of 29 indicators, and uses a 1 km × 1 km grid as the evaluation unit, combined with AHP-EM for combined weighting, aiming to more comprehensively and precisely illustrate the spatiotemporal patterns of ecological vulnerability in the study area. To address the aforementioned gaps, this study systematically conducts spatiotemporal pattern identification, driver mechanism analysis, and future scenario prediction of ecological vulnerability. The specific goals include: (1) revealing the spatial and temporal patterns and evolution mechanisms of ecological vulnerability in the YRSR under the “warm and humid” condition; (2) identifying key drivers and quantifying their contributions to the evolution of vulnerability; (3) and simulating future ecological risks under multiple scenarios and constructing differentiated collaborative governance strategies. By achieving these objectives, this study aims to provide scientific support for the sustainable development of the YRSR and similar alpine basins, directly serving global agendas including SDG 13, SDG 15, SDG 6, and SDG 11 on resilient community building.

2. Methods

This study builds a multi-model framework for ecological vulnerability assessment based on a four-stage workflow: “data foundation–vulnerability characterization–mechanism decomposition–scenario prediction”. The framework follows the progressive logic of “diagnosis–attribution–prediction”, aligning with the full cognitive chain from state identification to mechanism analysis and future prediction, thereby overcoming the limitations of single-method approaches.

First, standardized geographic data and a detailed index system provide the empirical basis for quantifying current vulnerability. As weight allocation critically affects assessment credibility, a single weighting method has clear drawbacks: principal component analysis may mask ecological differences, the analytic hierarchy process is susceptible to expert bias, and the entropy weight method, though objective, may ignore policy relevance [17]. Hence, a fused weighting strategy combining subjective and objective information is necessary. The AHP–entropy weight method retains expert knowledge while enhancing result robustness, making it more suitable for mountainous regions with uneven data quality and significant spatial heterogeneity.

Second, XGBoost-SHAP decomposes underlying driving factors, revealing non-additive and threshold-dependent effects, and shifts analysis from descriptive mapping to mechanistic understanding. XGBoost is chosen over linear regression or random forests due to the complex nonlinear interactions among driving factors in the YRSR. The gradient boosting algorithm captures non-monotonic relationships and interaction effects among temperature, precipitation, terrain, and human pressure, while SHAP values ensure interpretability of the “black-box” model—addressing the lack of interpretability in neural networks and deep learning models for geographical attribution. This combination outperforms statistical methods like Geodetector in ecological vulnerability research by not predetermining factor interaction forms, making it more suitable for exploratory studies with unclear driving mechanisms.

Third, the GA-PLUS land use simulation model predicts future spatial configurations under different policy scenarios, shifting assessment from retrospective diagnosis to forward-looking risk prediction. The YRSR features fragmented topography and significant spatial clustering of human activities. GA’s global optimization effectively handles spatial competition under complex plateau terrain constraints, while PLUS’s patch generation strategy better reflects the “point-axis” diffusion of human activities in mountainous areas. Their coupling ensures both optimal macro-level quantity structure and reasonable micro-level spatial patch layout, offering greater regional adaptability than standard cellular automata models in simulating refined land use evolution in mountainous regions.

These three modules are not simply combined but form a complementary methodology: the PSR-AHP–entropy weight module provides a static quantitative benchmark for vulnerability states, the XGBoost-SHAP module analyzes dynamic driving mechanisms, and the GA-PLUS module enables spatially explicit future scenario prediction. The output of each module serves as input constraints for the next, forming a closed-loop “state–mechanism–scenario” analysis. This enables assessment results with spatial accuracy, mechanistic depth, and predictive validity, better supporting differentiated management decisions in the YRSR than single-model assessments.

2.1. Data Standardization

For comparison and analysis on a uniform scale, standardization of the raw data is required to eliminate unit-induced discrepancies [18]. The process is as follows:

Positive indicators:

$$X_{ij}^{\prime }={\displaystyle \frac{(x_{ij}-min x_j)}{(max x_j-min x_j)}}$$

(1)

Negative indicators:

$$X_{ij}^{\prime }={\displaystyle \frac{(max x_j-x_{ij})}{(max x_j-min x_j)}}$$

(2)

where X′_ij denotes the normalized value of the j-th indicator for year i, while x_ij refers to its raw value [19]. The terms min x_j and max x_j represent the smallest and largest observed values of the j-th indicator over the entire time series.

2.2. Evaluation Index System

The framework of indicators was established based on the PSR model, covering three dimensions: pressure on resources and the environment, their quality, and the capacity for protection and restoration (

Table 1. Ecological vulnerability index system.

Quasi-Measurement Layer	Indicators (Code)	Attribute	W
Pressure	Landscape disturbance index (Q1)	+	0.0416
	Soil erodibility (Q2)	+	0.0539
	Potential evapotranspiration (Q3)	+	0.0363
	Nighttime lights (Q4)	+	0.0275
	Population density (Q5)	+	0.0106
	GDP (Q6)	+	0.0089
	Agricultural land area ratio (Q7)	+	0.0063
Status	Elevation (Q8)	+	0.0148
	Slop (Q9)	+	0.0105
	Topographic relief (Q10)	+	0.0085
	Surface temperature (Q11)	+	0.0271
	Soil type (Q12)	−	0.2187
	Land use type (Q13)	−	0.0497
	Water network density (Q14)	−	0.1438
	Annual precipitation (Q15)	−	0.0105
	Annual average temperature (Q16)	−	0.0035
	Annual evaporation (Q17)	+	0.1484
	Annual average wind speed (Q18)	+	0.0014
	Annual sunshine duration (Q19)	+	0.0019
	Habitat quality (Q20)	−	0.0103
	Water yield (Q21)	−	0.0077
	Carbon storage (Q22)	−	0.0058
Response	NPP (Q23)	−	0.0191
	NDVI (Q24)	−	0.0172
	Vegetation coverage (Q25)	−	0.0100
	Enhanced vegetation index (Q26)	−	0.0128
	Ecological protection red line (Q27)	−	0.0811
	Residents’ education level (Q28)	−	0.0072
	Ecological compensation fund (Q29)	−	0.0057

“+” indicates positive, “−” indicates negative.

).

2.3. Calculation of Weights

To combine subjective and objective weights, we use a linear weighting method [18]. Let the vector of subjective weights be W_S and the vector of objective weights be W_O, then the combined weight W_C is:

$$W_c=\alpha W_s+(1-\alpha )W_o$$

(3)

where α (0 ≤ α ≤ 1) balances expert experience and data information. To avoid arbitrariness in setting α we adopt an equal-weighted scheme (α = β = 0.5).

2.4. Ecological Vulnerability Index

The weighted sum of the evaluation scores of each indicator is calculated to obtain the comprehensive ecological vulnerability index [20].

$$EVI={\sum }_{i=1}^n{x}_{ij}\times W_i$$

(4)

where W_i represents the weight for indicator i, x_ij denotes the normalized value, and EVI is the ecological vulnerability index.

2.5. XGBoost-SHAP

XGBoost builds upon gradient boosting by sequentially fitting decision trees to residuals, progressively minimizing prediction error [21]. The formula is as follows:

$$\widehat{y_i}={\displaystyle \sum _{k=1}^k}{f}_k\left(x_i\right), f_k\in F(i=\mathrm{1,2},\dots ,n)$$

(5)

where ŷ_i represents the final predicted value of the i-th sample; x_i is the feature vector of the i-th sample; k denotes the total number of decision trees trained in the model; F represents the function space composed of all possible regression trees (also referred to as “the set of trees”); and n is the total number of samples [22].

Based on cooperative game theory, SHAP attributes to each input variable a contribution score reflecting its marginal impact on individual predictions—distinguishing both magnitude and direction of influence [23]. The calculation formula for the Shapley value is:

$$\oslash =\sum _{S\subseteq N\left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}}\left(v(S\cup \left\{i\right\})-v\left(S\right)\right)$$

(6)

where N denotes the complete set of features; S is a subset of N, |S| indicates its cardinality; v(S) quantifies the model’s prediction performance when trained or evaluated using only the features in |S| [24]; and v(S ⋃ {i}) represents the corresponding performance metric when feature i is added to |S|.

2.6. GA-Plus

Drawing on regional policy documents, ecological planning targets, and prior scenario frameworks [25,26], this study defines four distinct development pathways for the YRSR: “Natural Development” (ND), representing unregulated land use evolution; “Ecological Priority” (EP), prioritizing ecosystem service enhancement; “Economic Development” (ED), emphasizing GDP-oriented land allocation; and “Coordinated Coevolution” (CD), integrating ecological sustainability with socioeconomic advancement [27,28]. A genetic algorithm (GA) with enhanced search stability—leveraging high-dimensional parameter encoding and adaptive mutation strategies [29]—is employed to solve these complex, multi-objective problems and generate spatially explicit 2050 land use allocations. The specifics of each scenario are detailed below:

ND: This scenario follows the historical development trend of the study area. No restrictions are imposed on land type conversion, and no influence from policies or planning is considered. Based on land use transformation patterns from 2010 to 2020, the Markov model is used to simulate land use conditions in 2050.

EP: With the goal of maximizing ecological value, this scenario solves for the acreage of each land use category in 2050. The total ecosystem service value of each land category is extracted and used as the coefficient corresponding to the ecological benefit objective function [30]. The ecological benefit function max h(x) is set as follows:

$$max h(x)=\mathrm{m}\mathrm{a}\mathrm{x}{\sum }_{j=1}^6{e}_j{\sum }_{i=1}^6{x_{ij}S}_i$$

(7)

where h(x) represents the ecological benefit values of the YRSR (in ten thousand yuan), x_ij denotes the probability that the i-th land use type in 2020 is transformed into the j-th land use type in 2050, e_j is the ecological the value of land type j (in ten thousand yuan/km²), and S_i is the area of each land type in 2020.

ED: To maximize economic benefits, this scenario determines the extent of various land use categories in 2050. The spatial GDP distribution data for each land category is extracted and used as coefficients for the economic benefit objective function. The objective function for economic benefit max g(x) is set as follows:

$$max g(x)=\mathrm{m}\mathrm{a}\mathrm{x}{\sum }_{j=1}^6{c}_j{\sum }_{i=1}^6{x}_{ij}{S}_i$$

(8)

where g(x) represents the economic development benefit function of the YRSR (in ten thousand yuan), c_j is the economic benefit derived from the j-th type of land category (in ten thousand yuan/km²), and S_i is the spatial extent of each type of land category in 2020.

CD: In this scenario, the linear weighted summation method is applied. Ecological conservation and economic growth are coupled and aligned as joint development goals [31,32]. Based on the relevant literature, w₁ is set to 0.5, a₁ is set to 0.9023, w₂ to 0.5, and a₂ to 0.0988. The coupling and coordination objective function max f(x) is set as follows:

$$max f\left(x\right)=w_1\times a_1\times h\left(x\right)+w_2\times a_2\times g\left(x\right)$$

(9)

where f(x) is the coupling coordination comprehensive function (in ten thousand yuan), and S_i denotes the coverage of each land category in 2020.

3. Study Area and Data Sources

3.1. Study Area

The YRSR is located in the northeastern part of the Qinghai–Tibet Plateau (31°29′–36°23′ N, 96°50′–103°39′ E), which is known as the “Chinese Water Tower” (Figure 1). It covers roughly 132,000 km², with an average altitude of about 4300 m. This area has a fragile alpine ecosystem dominated by high-altitude meadows and high-altitude grasslands. Its hydrological system is complex and diverse, including a winding main river course of about 1553 km and widely distributed wetlands, lakes, marshes, and grasslands. These hydrological elements play an important role in water conservation, runoff regulation, and biodiversity protection. The terrain is mainly grassland, accounting for more than 80% of the vegetation, with scattered forests, farmlands, and human settlements. The average population density of this area is less than 0.5 people/km². With the dual effects of climate warming and local overgrazing, this area is facing increasingly severe ecological and environmental pressures.

3.2. Data Sources

This paper uses data covering land use from 2000 to 2020, NDVI, meteorological, soil, DEM, etc. Among them, land use and NDVI data are sourced from the China Academy of Sciences’ Resource and Environmental Science Data Center http://www.resdc.cn (accessed on 5 June 2025). Precipitation, temperature, potential evapotranspiration, etc., are obtained from the National Tibetan Plateau Data Center. https://data.tpdc.ac.cn (accessed on 12 October 2025). Soil data come from the Cold and Arid Region Scientific Big Data Center http://bdc.casnw.net/index.shtml (accessed on 3 September 2025). DEM data were acquired from the Geospatial Data Cloud http://www.gscloud.cn (accessed on 24 September 2025). CMIP6 data are acquired from WorldClim https://worldclim.org/ (accessed on 14 November 2025) and resampled to a 1 km grid cell size, with the projection spatial reference uniformly set to WGS_1984_Albers.

4. Results

4.1. Changes in Natural and Socioeconomic Factors

4.1.1. The Trend of “Warm and Humid Climate”

A two-decade assessment of climatic shifts in the Yellow River Source Region reveals distinct spatial patterns in temperature and precipitation dynamics. Precipitation exhibits a pronounced southward intensification trend—rising markedly in southern and especially southeastern subregions, while remaining largely unchanged across northern areas (Figure 2). This gradient strengthens progressively from north to south. Concurrently, regional warming has outpaced the global average, registering an overall mean increase of 0.336 °C/10a. However, warming is spatially uneven: the fastest rates occur in the central and northern zones—particularly within river valleys and lowland basins—whereas the western and southeastern high-elevation mountainous areas experience comparatively muted warming, underscoring topographic modulation of thermal change.

The “warming–humidifying” climate shift varies significantly across space. Combining the precipitation pattern (“south-up, north-stable”) with the temperature pattern (“central/north-fast, west/southeast-slow”) defines three response zones: In central–southern areas, strong precipitation increase combined with moderate warming produces the strongest warming–humidification signal. In northern zones, modest precipitation gain combined with moderate warming produces a weaker humidification signal. In western highlands, minimal changes in both precipitation and temperature result in the slowest warming–humidification.

4.1.2. Key Ecological Process Responses

Water yield rose in two phases: an increase of 19.1% between 2000 and 2010, followed by an increase of 56.5% between 2010 and 2020, indicating a clear acceleration (Figure 3). Over the period from 2000 to 2020, cumulative precipitation increased by 132.2 mm, which corresponds to a rise of 23.6%, confirming sustained and slightly accelerating growth, consistent with water yield trends. Statistical analysis shows a strong linear link between precipitation and water yield: each 1 mm increase in precipitation leads to a 0.687 mm increase in water yield, reflecting high runoff efficiency. Thus, rising precipitation was the primary direct driver of water yield expansion. Temperature showed only a weak statistical association with water yield and acted indirectly through biophysical or hydrological feedbacks, not as a dominant direct control.

4.1.3. Ecosystem State Changes Based on NPP

This study analyzed the changing trends based on the NPP data from 2000 to 2020, and the main results are as follows (Figure 4): 82.06% of pixels showed an upward trend and 16.57% exhibited a downward trend, while stable pixels accounted for only 1.37%, presenting an overall pattern of “strong upward, weak stable, and low downward” changes. The spatial distribution of NPP exhibits a “southeast to northwest decreasing” pattern: the southeastern Rongga Plateau and the Taohe–Daxiahe River Basin show the highest baseline NPP values, while the northwestern Yangtze River Source Park has the lowest NPP due to high elevation and low precipitation. In terms of trend, non-significantly increasing areas dominate spatially, with very few stable regions, forming a spatial pattern characterized by “large-scale enhancement and small-scale stability.”

Climatic amelioration—specifically rising temperatures coupled with increased moisture availability—emerges as the principal temporal driver, synergistically improving growing-season hydrothermal conditions and thereby boosting vegetation productivity. Spatially, NPP dynamics align closely with anthropogenic influence and ecosystem integrity: robust gains concentrate in ecologically resilient, low-disturbance zones—including the upper reaches of Longyangxia Reservoir and the Taohe–Daxiahe catchments—where natural vegetation remains largely intact. Conversely, NPP declines are spatially fragmented yet disproportionately severe in locations subjected to intensive human activity or inherently vulnerable ecosystems—such as Maduo County, major road and rail corridors, the river-source-to-Maqu stretch, and the desertified Gonghe Basin. This distribution underscores a fundamental linkage: sustained NPP improvement is strongly contingent upon minimal anthropogenic interference and the preservation of robust ecological foundations.

4.1.4. The Coupling Effect of Natural and Anthropogenic Pressures

Between 2000 and 2020, the spatial distribution of pressure indicators remained highly consistent (Figure 5). Regions characterized by high population and GDP density experienced more pronounced landscape disturbance and moderate soil erosion. Temporally, the first decade (2000–2010) was marked by pressure accumulation, whereas the subsequent period (2010–2020) saw the diffusion and intensification of pressures, particularly in urban fringe zones and along transport corridors, where multiple pressures intersected. Throughout the study period, potential evapotranspiration and soil erosion in the YRSR remained relatively stable in both space and time. Human activity intensity remained low, with no significant expansion, providing the basis for sustained improvements in regional Net Primary Productivity: limited anthropogenic disturbance coupled with stable hydrothermal conditions facilitated widespread gains in vegetation productivity. However, localized soil erosion continued to constrain ecological restoration in some areas.

Spatially, the YRSR exhibits a persistent west–low and east–high gradient, alongside elevated vulnerability in the north and reduced levels in the south (

Table 2. Classification of vulnerability levels.

Vulnerability Level	Vulnerability Score	Level
Not vulnerable	<0.23	—
Extremely low vulnerability	0.23–0.37	I
Low vulnerability	0.37–0.46	II
Moderately vulnerable	0.46–0.56	III
Highly vulnerable	0.56–0.68	IV
Extremely high vulnerability	>0.68	V

), mirroring natural gradients and the spatial heterogeneity of human pressures (Figure 6).

Low-vulnerability zones lie mainly in the western sector and along the southeastern Zoige Wetland, benefiting from favorable hydrothermal conditions, low soil erosion, strong vegetation productivity, and healthy NPP performance. Medium vulnerability dominates the central grassland belt, where NPP remains functional but suboptimal. High vulnerability appears as fragmented, patchy clusters across the central and southern YRSR, coinciding with areas where human stressors intersect with inherently fragile ecological conditions.

Temporally, vulnerability remained largely static from 2000 to 2010, characterized by subtle, two-way adjustments between neighboring classes and no net shift toward degradation or improvement—reflecting a relative equilibrium. From 2010 onward, a clear one-way intensification emerged, marked by a sharp rise in vulnerability and a transition from “stable fine-tuning” to “accelerated deterioration” (Figure 7). Thus, 2010 serves as a key turning point. After 2010, transitions between classes became stronger and more uneven; moves from high to extremely high vulnerability turned nearly irreversible, while the extremely high class exhibited remarkable persistence. Concurrently, low-vulnerability areas shrank, whereas medium, high, and especially extremely high vulnerability zones expanded, driving an overall rise in regional vulnerability.

4.2. Ecological Vulnerability and Its Driving Factors

4.2.1. Spatiotemporal Differentiation of Ecological Vulnerability

Robustness checks—varying the α parameter and introducing random data perturbations—confirmed that these spatial and temporal patterns persist under alternative specifications. The identified 2010 turning point and subsequent vulnerability intensification are not driven by arbitrary parameter choices or data noise. Specifically, the west–east and north–south spatial gradients, the pre-2010 stability, and the post-2010 one-way deterioration remained qualitatively unchanged across all tests, supporting the reliability of our main conclusions.

4.2.2. Identification of Main Driving Factors

Given the YRSR’s defining traits—high elevation, aridity, and heightened ecological fragility—a targeted correlation analysis was conducted between ecological vulnerability and key environmental and anthropogenic variables (Figure 8). Potential evapotranspiration (Q3) showed the strongest positive correlation with vulnerability (R² = 0.952), consistent with the region’s characteristic evaporation-dominated water deficit. Elevated evaporation is associated with increased near-surface desiccation, which may contribute to alpine meadow degradation and desertification—processes that, in turn, correlate with higher system susceptibility. Surface and air temperature also showed strong positive correlations, suggesting a potential role of warming trends. Rising temperatures may accelerate permafrost thaw and alter alpine plant community composition, which could be linked to reduced ecosystem resilience.

Among the negatively correlated factors, annual average wind speed (Q18) exhibited the strongest inverse relationship. This may indirectly influence vegetation phenology and growth rhythm, though the underlying mechanisms require further investigation. NPP (Q23) followed closely, reflecting its function as an indicator of vegetation vigor. Denser and healthier vegetation tends to be associated with improved soil moisture retention and surface stabilization—factors that correlate with lower vulnerability. Elevation (Q8) also displayed a significant negative correlation, likely reflecting topographic controls. Higher elevations are associated with limited human access and more intact permafrost layers, which tend to coincide with greater ecological stability. Conversely, lower-elevation valleys concentrate socioeconomic activity and exhibit intensified erosion, corresponding with markedly higher vulnerability.

Using the XGBoost-SHAP framework, a machine learning attribution analysis quantified the relative influence of drivers (Figure 9). Results show a “core-concentrated, peripherally diffuse” structure in vulnerability determinants.

Temperature (Q16) and potential evapotranspiration (Q3) together explain over 90% of spatial variance in vulnerability. Temperature alone contributes nearly half, affirming its role as the foundational thermal regulator governing vegetation phenology, soil moisture, and evaporation. Thus, temperature functions as the overarching driver shaping vulnerability patterns. Potential evapotranspiration and temperature work together to determine the regional water balance and affect vulnerability indicators such as drought and vegetation degradation. Apart from these two factors, the contribution rates of all remaining factors are below 5%. Their marginal effect declines sharply as more factors are added. These factors only provide supplementary effects for spatial differentiation of vulnerability.

4.3. Ecological Risk Prediction in Different Scenarios

4.3.1. Projection of Land Use Dynamics and Climate

An analysis of land cover dynamics from 2000 to 2020 according to transition matrices and area change metrics, reveals key features of ecosystem evolution in the YRSR (Figure 10). Grassland remains the dominant land-cover class throughout the period, with only minor fluctuations in extent; its structural persistence underscores the fundamental resilience and relative stability of the alpine grassland system. Notably, net conversions from unused land into grassland have occurred consistently, signaling tangible success of ecological restoration initiatives, including livestock reduction programs, grassland recovery efforts, and wetland conservation measures in reinforcing the region’s ecological backbone. However, bidirectional exchanges between grassland and unused land remain frequent, highlighting the system’s heightened responsiveness to both climatic variability and localized anthropogenic pressures—suggesting that extreme weather events or unsustainable land management may still trigger reversible or irreversible grassland degradation.

Under scenario-based projections (Figure 11), land use trajectories diverge markedly across policy pathways. ND: Land-cover shifts follow historical trends with minimal deviation, reflecting inertia rather than intentional steering. EP: Substantial gains in forest cover, grassland extent, and water-body area, achieved largely through farmland-to-ecological-land conversion; construction land expansion is tightly curbed, yielding demonstrable ecosystem enhancement. ED: Aggressive conversion of farmland, forest, and grassland into built-up areas, accompanied by widespread reclamation of marginal lands, placing acute pressure on natural capital and biodiversity.

CD: A deliberate balance—construction land growth constrained, with new development preferentially allocated to previously disturbed or unused land; farmland and ecologically critical zones actively safeguarded, enabling synergistic progress toward both developmental goals and environmental sustainability.

Under the three SSP scenarios, temperature in the YRSR is higher in the east and lower in the west (Figure 12). This east–west gradient—warmer in the southeast and cooler in the northwest—mirrors topographic controls. Lower-elevation southeastern valleys receive warm, moist air masses, while higher-elevation northwestern sectors remain under cold, stable plateau circulation.

From SSP126 to SSP585, warming escalates progressively in both spatial extent and intensity. SSP585 projects the strongest warming, especially across mid to low elevations. SSP126 retains a broader expanse of cooler areas, confirming the climate benefits of stringent emissions control. Precipitation follows a complementary southeast–northwest gradient: abundant in the southeast and scarce in the northwest. Patterns are broadly similar across scenarios, but subtle differences emerge. Under SSP585, the southeastern high-precipitation zone expands slightly, implying a higher likelihood of intense rainfall and related hazards. Conversely, SSP126 yields a more homogeneous precipitation distribution. This indicates lower interannual variability and greater predictability in water availability, supporting ecosystem stability and adaptive capacity in this water-limited alpine region.

4.3.2. Future Ecological Vulnerability Response

Across the four policy scenarios, regional average ecological vulnerability ranges narrowly between 0.414 and 0.419, all falling within the moderate category. The proportion of “extremely high-vulnerability” zones remains constant at 2.3%, suggesting that core high-risk areas exhibit limited sensitivity to policy variation (Figure 13). These high-risk zones are predominantly clustered in the northeast quadrant—an area that consistently shows elevated vulnerability. In contrast, low-vulnerability areas are located in the western and southern sectors, where dense vegetation, gentle topography, and strong ecosystem resilience are observed.

A comparative assessment indicates measurable differences across scenarios. Under EP, the average vulnerability shows a slight increase, accompanied by a minor spatial expansion of degraded patches in the northeast. Under ED, the average vulnerability is the lowest among the four scenarios, and high-risk clusters display noticeably lower color intensity. Under CD, vulnerability levels are nearly identical to those under ED.

To systematically evaluate policy performance, vulnerability changes were benchmarked against the ND baseline (Figure 14). ED reduces regional average vulnerability by 0.001, confirming localized mitigation gains. However, positive vulnerability differences (VD > 0) appear sporadically, indicating uneven returns on conservation investment. EP increases average vulnerability by 0.006, with elevated risk concentrated in the northeast—signaling development-driven ecological deterioration. Under CD, mean vulnerability shows negligible net change relative to ND. However, localized “hotspots” and “cold spots” coexist, suggesting context-dependent outcomes: ecological gains in protected zones are offset by residual pressures in growth corridors. Notably, the “CD versus ED” comparison shows virtually no difference in overall vulnerability, with only minor local deviations. This indicates that coordinated development can achieve ecological outcomes comparable to strict protection without sacrificing economic objectives, making CD a pragmatically balanced governance strategy.

5. Discussion

5.1. Climate Change Differentiation

Most existing regional and global studies have primarily emphasized the overarching “warming and wetting” trend of the Qinghai–Tibet Plateau. However, they fall short in precisely capturing the decoupling between temperature and precipitation changes in terms of magnitude and spatial patterns. This research identifies distinct zones of “strong”, “weak”, and “flat” warming–wetting variations within the YRSR. These findings directly support SDG 13 by moving beyond averaged trends to reveal localized climate risks [33].

Notably, high temperatures are primarily concentrated in the central and northern river valleys. Yet these areas do not fully coincide with zones showing substantial precipitation increases. This finding aligns with the precipitation pattern characterized as “high south, low north” in the northeastern plateau, as reported by [34]. Such spatial mismatches suggest that thermal and hydrological changes may diverge under the influence of complex terrain and local atmospheric circulation, thereby shaping diverse regional configurations. This spatial variability implies that ecological assessments based solely on averaged climatic trends may overlook critical localized risks. In the northern “strong warming but weak wetting” zone, for example, heightened warming intensifies potential evapotranspiration. This could worsen water scarcity and lead to a distinct pattern of ecological decline. Addressing such risks is essential for achieving SDG 13 targets related to climate resilience and adaptation.

Corresponding changes in water yield further reflect this spatial divergence. In the southern and especially southeastern areas, where precipitation increases are most pronounced, water yield functions have notably strengthened. The central-south region remains precipitation-driven, with moderate warming further facilitating runoff generation. In the north, increased evaporative demand largely offsets precipitation gains. This results in stagnant or negative water yield responses and greater uncertainty in water supply reliability. The western part displays a moderated warming–wetting trend with comparatively stable water yield, though its fragile ecosystem remains sensitive to climatic shocks.

Consequently, a unidimensional “warming and wetting” metric is insufficient for reliably predicting ecological outcomes. It is necessary to distinguish between “wet-dominated” and “warm-dominated” climate regimes. It is also important to incorporate contextual variables such as topography, vegetation cover, and soil characteristics. This will help systematically unravel the mechanisms shaping water yield dynamics and the evolution of ecological vulnerability. Such mechanistic understanding is crucial for designing informed climate actions under SDG 13, particularly in high-altitude, data-sparse regions.

5.2. Evolution of Ecological Vulnerability

The spatial pattern of ecological vulnerability arises from interactions among natural constraints, human pressures, and ecosystem response capacity. Natural conditions provide the static basis: contrasts between the southeast and northwest predetermine regional sensitivity and recovery potential [35]. Understanding this pattern supports SDG 15, particularly target 15.3 on land restoration [36].

Human activity acts as a key external force that worsens vulnerability. From 2000 to 2020, pressures along urban peripheries and transport corridors expanded disturbances from points to broader areas, creating degradation “hotspots” even in the relatively favorable southeast. This finding directly informs territorial planning under SDG 11, especially target 11.3 on sustainable urbanization [37]. Despite localized pressures, regional NPP showed an overall upward trend. This paradox reveals that low human activity intensity, absence of drastic expansion, and stable natural conditions have created buffering capacity for ecosystem recovery. Consequently, high-vulnerability areas represent the spatial overlap of “adverse natural backgrounds” with “persistent anthropogenic pressures.”

Crucially, climate conditions—particularly thermal and moisture regimes—fundamentally shape vulnerability patterns and trajectories. Therefore, ecological restoration and adaptive management must prioritize climate change mitigation and alleviation of warming-induced drought stress. Governance that narrowly regulates local pressures without addressing this climatic driver will yield limited outcomes. This insight is central to climate adaptation strategies for high-altitude ecosystems. Differentiated management strategies are essential. In areas with high vulnerability driven by temperature and evapotranspiration, efforts should focus on strengthening climate resilience. In degradation hotspots shaped by anthropogenic pressures, land use regulation and development intensity must be tightened. These measures support SDG 11.6 (reducing cities’ environmental impact) and should be embedded within a climate adaptation framework [38].

Finally, “marginal factors” matter. Although low-contribution edge factors do not govern the overall pattern, their local effects can be critical. While adhering to the “climate dominance” principle, refined grid-based management must account for combined local effects. Such fine-scale governance enhances alpine ecosystem resilience and offers a transferable framework for other high-altitude regions facing similar pressures.

5.3. Future of Ecological Risks

Simulation results show that mean ecological vulnerability varies only slightly across the four scenarios, but the spatial distribution of high-risk areas remains highly stable. The proportion of core risk zones is consistent across scenarios. This finding aligns with [39], suggesting that natural background rigidity imposes a threshold or inertia effect on human intervention. This directly supports SDG 15 (Life on Land), particularly target 15.3 on land restoration.

Ecological risk outcomes differ markedly among pathways. Under EP, regional mean vulnerability increases and high-risk areas in the northeast expand. This illustrates how development can amplify vulnerability in settings with weak ecological constraints [40]. Under ED, mean vulnerability is lowest and high-risk zones shrink, suggesting that spatially targeted policies can alleviate degradation pressures [41]. Under CD, vulnerability levels are comparable to ED, yet with an intricate interlocking pattern of high- and low-risk areas. Through scientific territorial planning and stringent ecological controls, CD maintains overall risk near protection-only levels while accommodating addressing future ecological risks requires a transition from uniform management to precise spatial governance [42,43]. Land use regulation should function as a rigid constraint that shapes human activity patterns and curbs high-intensity disturbances [44,45]. Ecosystem-based climate adaptation should strengthen natural system resilience. These two approaches must be integrated within a “protection priority” spatial framework. The CD pathway operationalizes this synergy by optimizing the spatial distribution of human activities.

A four-tier protection priority classification offers a clear framework. In northeastern extremely high-risk zones, the strictest protection and proactive adaptation are required, combining development restrictions with nature-based measures. In central transitional moderate-risk areas, flexible spatial governance should use positive-list controls and adaptive management. In western and southern low-risk zones, green development can proceed within a preventive adaptation framework, with revenues directed toward ecological restoration. This creates a virtuous cycle where development supports protection.

5.4. Limitations

Although this study utilized multi-source data and a 1 km resolution grid, some indicators still had issues of insufficient temporal and spatial resolution or were missing, which might affect the precise depiction of micro ecological processes. This study’s scenario simulation relied on a set of assumptions. Although it referred to the IPCC scenarios and local planning as much as possible, it still could not completely avoid the uncertainties of future social economic development, technological progress, and extreme climate events. Although this study included indicators of human activity pressure, it mainly focused on their spatial intensity, and paid insufficient attention to how social dimensions such as social economic structure, institutional arrangements, community adaptability, and cultural cognition affect vulnerability perception, behavioral responses, and the construction of resilience. Ignoring the social dimension might reduce the feasibility and sustainability of policy recommendations.

6. Conclusions

This study developed a framework to assess ecological vulnerability in the YRSR. It combined multi-source data within the PSR model, employing AHP-EM weighting and XGBoost-SHAP attribution analysis. The research systematically investigated spatiotemporal dynamics, dominant drivers, and projected risk scenarios. The results reveal that the ongoing “warming and humidification” trend exhibits spatial heterogeneity: southern enhancement with northern stability, and intensified warming in central and northern areas. This heterogeneity has led to divergent responses in key ecological processes, such as water yield. Thus, relying solely on aggregate “warming and humidification” metrics is inadequate for vulnerability assessment. This finding directly supports SDG 13 by highlighting the need for localized climate risk assessments. The spatial distribution of vulnerability is governed by a composite mechanism. Climatic factors are the primary drivers, while human activities play a modulating role. Temperature and potential evapotranspiration together contribute over 90% to observed vulnerability. This confirms that thermal and moisture regimes are fundamental controls in this high-altitude environment, underscoring the relevance to SDG 15, particularly target 15.3 on land degradation and restoration. Multi-scenario modeling consistently identifies the northeastern sector as a persistently high-risk zone. Simulations further indicate that a coordinated development pathway can effectively reconcile ecological integrity with developmental demands. This offers quantifiable insights for policymaking that align with SDG 11, especially target 11.3 on sustainable urbanization and land use planning. Based on these findings, the study establishes a four-tiered priority protection framework. It also proposes a spatial governance strategy encompassing “strict preservation, flexible regulation, and controlled utilization.” Together, they provide a theoretical foundation for targeted ecological conservation and adaptive management in the YRSR, reinforcing the study’s broader social and policy relevance.