Assessing the Climate and Land Use Impacts on Water Yield in the Upper Yellow River Basin: A Forest-Urbanizing Ecological Hotspot

Abstract

Understanding the drivers of water yield (WY) changes in ecologically sensitive, data-scarce watersheds is crucial for sustainable management, particularly in the context of accelerating forest expansion and urbanization. This study focuses on the upper Yellow River Basin (UYRB), a critical headwater region that supplies 60% of the Yellow River’s flow and is undergoing rapid land use transitions from 1990 to 2100. Using the Integrated Valuation of Ecosystem Services and Tradeoffs (InVEST) model and the Future Land-Use Simulation (FLUS) model, we quantify historical (1990–2020) and projected (2025–2100) WY dynamics under three SSP scenarios (SSP126, SSP370, and SSP585). InVEST, a spatially explicit ecohydrological model based on the Budyko framework, estimates WY by balancing precipitation and evapotranspiration. The FLUS model combines cellular automata (CA) with an artificial neural network (ANN)-based suitability evaluation and Markov chain-derived transition probabilities to simulate land-use change under multiple scenarios. Results show that WY increased significantly during the historical period (1990–2020), primarily driven by increased precipitation, with climate change accounting for 94% and land-use change for 6% of the total variation in WY. Under future scenarios (SSP126, SSP370, and SSP585), WY is projected to increase to 217 mm, 206 mm, and 201 mm, respectively. Meanwhile, the influence of land-use change is expected to diminish, with its contribution decreasing to 9.1%, 5.7%, and 3.1% under SSP126, SSP370, and SSP585, respectively. This decrease reflects the increasing strength of climate signals (especially extreme precipitation and evaporative demand), which masks the hydrological impacts of land-use transitions. These findings highlight the dominant role of climate change, the scenario-dependent effects of land-use change, and the urgent need for integrated climate–land management strategies in forest-urbanizing watersheds.

1. Introduction

Water yield (WY), an essential indicator of hydrological ecosystem services, is defined as the net difference between precipitation and actual evapotranspiration, representing the upper limit of water resources available to natural ecosystems and human societies [1,2]. Changes in WY, driven by climate change and human activities, are crucial for the development of ecosystems and human societies. With increasing global warming and human intervention, scientific research on the quantitative assessment and spatial distribution of water yield in ecosystems is essential to ensure the sustainable use of water resources, adapt to the impacts of climate change, and support stable socioeconomic development.

Climate change and land-use change, the main drivers of watershed hydrological processes, collectively influence the availability of water resources and the patterns of WY in watersheds worldwide [3,4,5,6,7]. Climate change directly impacts WY through changes in precipitation patterns, modifications in evapotranspiration processes, and temperature shifts [8,9,10]. In contrast, land-use change primarily affects the redistribution mechanisms of water resources, such as infiltration, evapotranspiration, and soil moisture storage, thereby influencing the availability and effectiveness of water resources [11,12,13,14,15]. Many studies have explored the individual or combined impacts of climate change and land-use change on WY over historical periods [3,16,17]. However, there is a lack of studies that systematically assess how the combined effects of long-term climate and land-use changes jointly influence WY dynamics, particularly in high-altitude basins where hydrological responses are more sensitive and complex.

To distinguish the relative impacts of climate change and land use on water yield (WY), three primary approaches are commonly used: paired catchment analysis, statistical attribution, and numerical modeling [8,18,19]. Of these, modeling approaches are particularly effective for evaluating historical and future WY changes under diverse scenarios [4,13]. Hydrological models can be broadly classified into two categories based on their complexity and data demands. Physically based models, such as SWAT and VIC, simulate detailed processes across the hydrological cycle, including infiltration, percolation, runoff, and evapotranspiration [20,21]. However, they are parameter-intensive and often face equifinality issues, where different parameter sets yield similar results but differ in physical realism. Calibration becomes particularly challenging in data-scarce basins, such as the UYRB, prompting a growing reliance on remote-sensing products (e.g., ET, LAI, soil moisture) for auxiliary constraints. In contrast, ecosystem service models—such as ARIES [22] and InVEST [23,24]—adopt simplified hydrological representations, often using annual water balance or Budyko-type formulations. InVEST estimates WY as the difference between precipitation and actual evapotranspiration, requiring minimal inputs (e.g., land cover, soil depth, PET). A significant advantage is its low calibration burden; for example, the InVEST model includes only one tunable parameter (Z), enhancing its usability in data-limited environments [25,26]. In addition, unlike many physically based models that operate at the sub-basin level, InVEST performs simulations at the raster (grid cell) scale, allowing a spatially explicit integration with land use/land cover maps. This enables direct coupling with land-use transition scenarios to evaluate the spatially heterogeneous impacts of LUCC on WY. However, this simplicity may overlook important processes such as groundwater–surface water interactions [27]. Overall, while traditional hydrological models offer higher physical fidelity, simplified ecosystem models, such as InVEST, are better suited for scenario-based planning in data-constrained regions [28].

The Yellow River Basin is considered the mother river of China, as its water serves 12% of the population, irrigates 17% of the arable land, and contributes to 14% of the nation’s GDP [29]. The upper Yellow River Basin (UYRB), accounting for 60% of the Yellow River Basin’s total water volume [30], has experienced significant climate warming and land-use changes [31,32,33,34], profoundly impacting the availability of water resources in the basin. Therefore, quantitatively evaluating the influence of climate and land-use changes on WY in the UYRB has become a pressing scientific challenge. Several researchers have studied the variations in runoff and the primary factors affecting runoff in the UYRB [30,35,36]. However, the findings of these studies are not entirely consistent, which could be attributed to differences in research scales, analytical approaches, and the extent of climatic or land-use changes. For instance, Lv et al. [37] found that climate change, especially increased precipitation, played a dominant role in recent runoff increases in the UYRB based on SWAT modeling and scenario simulations. In contrast, Su et al. [38] argued that in certain tributaries such as the Huangshui River, human activities—particularly reservoir operations and land-use changes—had a more significant impact on runoff reduction, especially during dry seasons. Existing studies have primarily focused on analyzing the effects of climate change and land-use change on runoff, with fewer studies investigating WY in the UYRB. Furthermore, these studies have not adequately addressed long-term changes in WY or explored future changes under different scenarios.

To address the limited understanding of long-term water-yield (WY) dynamics under integrated climate and land-use scenarios in high-altitude regions, this study focuses on the upper Yellow River Basin (UYRB), a climate-sensitive and ecologically critical headwater area. We propose a scenario-based assessment framework combining the FLUS model for land-use simulation, the InVEST model for WY estimation, and a suite of analytical tools including trend analysis, sensitivity analysis, and attribution analysis. The specific objectives of this study are as follows:

(1)

To evaluate the spatiotemporal trends and variability of WY in the UYRB over the historical period (1990–2020);

(2)

To disentangle and quantify the relative contributions of climate change and land-use change to historical WY variations using both sensitivity and scenario-based attribution methods;

(3)

To assess future WY changes (2025–2100) under combined climate (SSP126, SSP370, SSP585) and land-use scenarios, thereby filling the research gap in understanding long-term WY responses to integrated drivers in alpine basins.

2. Materials and Methods

2.1. Study Area

The upper Yellow River Basin (UYRB), located upstream of the Lanzhou hydrological station, spans 95°52′ E–104°12′ E and 32°09′ N–38°17′ N. This region covers an area of 223,000 km², accounting for approximately 28% of the entire Yellow River Basin [30,36]. Positioned as a transition zone between the northeastern Tibetan Plateau and the Loess Plateau, the UYRB features diverse topography, with elevations ranging from 1514 m to 6274 m and an average elevation of 3634 m. The landscape, dominated by mountains and plateaus, encompasses the Qilian Mountains and is intersected by several river systems, resulting in a complex geomorphic environment. The climate is characterized by cold, dry winters and hot, rainy summers, with approximately 500 mm of annual precipitation occurring mainly during the summer [39]. As a critical runoff source, the UYRB contributes around 60% of the Yellow River’s annual runoff [36], with precipitation being the primary contributor [40]. The study area’s varied soils and vegetation, including grasslands and forests, significantly influence the spatial and temporal dynamics of WY. To facilitate analysis, the UYRB was partitioned into six sub-basins based on the three-tier water-resources zoning framework implemented by the People’s Republic of China, as shown in Figure 1.

2.2. Data Sets

The datasets comprise historical and future climate projections, historical and future land use data, topographic characteristics (elevation and slope), population density distributions, socioeconomic indicators (GDP), soil properties, and naturalized runoff records (

Table 1. Datasets used in this study.

Data Type	Period	Source	Application
Precipitation, Tmax, Tmin	1990–2020	National Meteorological Information Centre (http://data.cma.cn/ (accessed on 6 August 2025))	Calculate potential evapotranspiration as input for the InVEST model.
Future climate (Precipitation, Tmax, Tmin)	2025–2100 (5-year step)	ISIMIP3b (GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, and MRI-ESM2-0)	Calculate future potential evapotranspiration as input for the InVEST model.
Historical land use	1990–2020	Annual land-cover product by [41]	As input for the InVEST model.
Future land use (LUH2)	2020–2100	LUH2 dataset (https://luh.umd.edu (accessed on 6 August 2025))	As input for the InVEST model.
Digital elevation model (DEM)	Static	NASA DEM	As input for the FLUS model.
Soil properties	Static	Harmonized World Soil Database (HWSD v1.2)	Includes soil depth, texture, used in the InVEST model.
Naturalized runoff	1998–2020	Yellow River Water Resources Bulletin (http://www.yrcc.gov.cn (accessed on 6 August 2025))	Used for InVEST model validation.
Population density	2010, 2015, 2020	WorldPop (https://hub.worldpop.org (accessed on 6 August 2025))	As input for the FLUS model.
GDP	2010, 2015, 2020	Product by [42]	As input for the FLUS model.
Remote sensing ET (GLEAM)	1990–2020	GLEAM v4.1a (https://www.gleam.eu (accessed on 6 August 2025))	Verify the InVEST model.

). The following section presents a comprehensive description of each dataset.

2.2.1. Climate Data

Historical daily precipitation, maximum temperature, and minimum temperature data (1990–2020) were obtained from over 2400 stations nationwide. These data were sourced from the National Meteorological Information Centre (http://data.cma.cn/ (accessed on 6 August 2025)) and subjected to rigorous quality control by the China Meteorological Administration (CMA). The data were interpolated using the ANUSPLIN v4.4 software [43], which was selected for its ability to incorporate elevation as a covariate—an essential feature for ensuring accuracy in regions with complex topography such as the upper Yellow River Basin (UYRB)—and restricted to the UYRB. Mean annual precipitation was calculated by aggregating daily precipitation data, while daily reference evapotranspiration was estimated using the Hargreaves equation, which is recommended for the InVEST model due to its reduced parameter dependence and reliable performance in data-scarce basins [4,44].

Future daily precipitation, maximum temperature, and minimum temperature data were obtained at five-year intervals from 2025 through 2100, based on the ensemble averages of four general circulation models (GCMs)—GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, and MRI-ESM2-0—under the ISIMIP3b framework. Bias correction and statistical downscaling were applied to CMIP6 model outputs using the ISIMIP3BASD v2.5.0 and W5E5 v2.0 methodologies. Three future scenarios were analyzed: SSP126 (sustainable development with low emissions), SSP370 (medium development), and SSP585 (rapid fossil fuel development with high emissions).

2.2.2. Land-Use Data

Land-use data for 1990–2020 were obtained from China’s 30 m annual land-cover product, created by [41]. This dataset is based on 5463 visually interpreted samples and has an overall accuracy of 80%. It was resampled to a 1 km resolution. The land-use data were categorized into six classes: cropland, grassland, forest land, water, unused land, and urban areas. These land use classes were defined based on GB/T21010-2007 [45] and further refined to align with regional land-cover characteristics observed in the UYRB.

The Land Use Harmonization Database (LUH2), part of the Coupled Model Intercomparison Project Phase 6 (CMIP6), provides annual land-use patterns from 850 to 2100 at a spatial resolution of 0.25° × 0.25°. To ensure compatibility with regional conditions, the LUH2 dataset was calibrated against historical land-use data [46]. First, discrepancies were identified between LUH2 and historical datasets regarding land-use categories and spatial extents in overlapping years, necessitating harmonization to achieve dataset consistency. Both datasets were therefore reclassified into six standardized land-use categories—grassland, cropland, forestland, water bodies, unused land, and urban areas. Since the LUH2 dataset does not reliably project future changes in water body extent, and previous studies have shown minimal impacts of water areas on watershed water yield [28], the water category was excluded from further analysis. Second, future land-use change rates were computed under three distinct scenarios based on the 2020 LUH2 projections. These rates were then applied to historical land-use data from 2020 to project future land-use demands for grassland, cropland, forestland, unused land, and urban areas.

2.2.3. Other Data

Elevation data (30 m resolution DEM) were collected from NASA, with derived slope data (in degrees) calculated using ArcGIS 10.8.2 software. Soil properties were sourced from the Harmonized World Soil Database (HWSD, URL: https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/ (accessed on 6 August 2025)). Naturalized river runoff records spanning 1998–2020 were extracted from the Lanzhou Station hydrological archives within the Yellow River Water Resources Bulletin (URL: http://www.yrcc.gov.cn (accessed on 6 August 2025)). Population density datasets for 2010, 2015, and 2020 were acquired from World Pop (URL: https://hub.worldpop.org/project/categories?id=18 (accessed on 6 August 2025)). Corresponding GDP data for the same years were obtained from [42], which provided accurate and corrected gross domestic product data. All socioeconomic data were resampled to a 1 km resolution using bilinear interpolation and temporally aligned with the land use data.

2.3. Methods

This study evaluates historical (1990–2020) and future (SSP126, SSP370, SSP585) water yield dynamics in the UYRB using the InVEST model with FLUS-projected land use and bias-corrected climate data, analyzing trends and drivers through statistical and scenario-based methods. The overall technical framework of this study is illustrated in Figure 2.

2.3.1. InVEST Water Yield Model

The InVEST water yield module utilizes the Budyko curve [47,48,49] and the annual average precipitation [46] (Equation (1)). It assumes all WY is combined and directed to the watershed outlet as runoff without distinguishing between surface water, groundwater, and baseflow [46]. For a detailed rationale, see the InVEST User’s Guide [50].

$$WY\left(x\right)=(1-{\displaystyle \frac{AET\left(x\right)}{P\left(x\right)}})\times P(x)$$

(1)

where WY (x), AET(x), and P(x) represent the annual water yield, the annual actual evapotranspiration, and the annual precipitation (mm) for each pixel, respectively.

The actual evapotranspiration (ET) of the three vegetation land-use types, grassland, forestland, and cropland, was calculated mainly using Equation (2).

$${\displaystyle \frac{AET\left(x\right)}{P\left(x\right)}}=1+{\displaystyle \frac{PET\left(x\right)}{P\left(x\right)}}-{[1+(PET\left(x\right)/{P\left(x\right))}^w]}^{1/w}$$

(2)

where PET(x) represents the potential evapotranspiration (mm) for each pixel, respectively; $w$ represents a catchment-specific ecohydrological parameter reflecting soil–climate interactions. According to the InVEST User’s Guide, the following equations were used to calculate PET(x) and $w$:

$$w=Z{\displaystyle \frac{AWC\left(x\right)}{P\left(x\right)}}+1.25$$

(3)

$$PET\left(x\right)=K_c\left(l_x\right)\times {ET}_0\left(x\right)$$

(4)

where Z is a seasonal constant, which needs to be obtained by comparing the simulated WY with the actual WY; $K_c\left(l_x\right)$ is the evapotranspiration coefficient of the plants (vegetation) associated with the pixel; ${ET}_0\left(x\right)$ is the reference evapotranspiration of the pixel; and $AWC\left(x\right)$ represents the amount of available soil water (mm). The calculation of $AWC\left(x\right)$ is referenced to the InVEST User’s Guide.

For unused land and urban areas, the actual evapotranspiration is directly calculated from the reference evapotranspiration, the upper limit of which is determined by precipitation, with the formula shown below:

$$AET\left(x\right)=Min\left(K_c{\left(l_x\right)ET}_0\right(x),P(x\left)\right)$$

(5)

where $K_c\left(l_x\right)$ is the evaporation factor for unused land and urban areas.

To validate the evapotranspiration (ET) estimates generated by the InVEST model in the absence of in situ ET observations, this study employed annual mean ET data derived from the remotely sensed GLEAM model, which estimates potential ET using the Penman equation. The GLEAM dataset was selected due to its extensive temporal coverage (1980 to present) and demonstrated accuracy in ET estimation. The most recent versions, GLEAM v4.1a and v4.1b, offer enhanced data quality and consistency. In particular, GLEAM v4.1a provides a global ET record spanning 44 years (1980–2023), integrating reanalysis-based radiation and air temperature, precipitation from gauge observations, reanalysis products, and satellite sources, as well as satellite-derived vegetation optical depth. Furthermore, a correction factor of 1.18, derived by [51] through a comparative analysis with flux tower observations in the Yellow River water conservation zone, including UYRB, was applied to the GLEAM ET estimates to improve their regional accuracy in this study.

2.3.2. FLUS Model

The FLUS model, described in detail in previous studies [52,53], was employed in this study to simulate the future spatial distribution of land use. Since the FLUS model typically simulates land-use changes at 5-year intervals [53], the observed transitions from 2015 to 2020 were used to generate a transition probability matrix via a Markov chain. Specifically, the Markov chain method quantifies the likelihood of land-use type conversions based on historical land-use data, and this transition probability matrix subsequently serves as a critical input for the cellular automata component of the FLUS model. The actual land-use data from 2020, along with the projected land-use demands under three scenarios (SSP126, SSP370, and SSP585) for the years 2025, 2030, 2035, 2040, 2045, 2050, 2055, 2060, 2065, 2070, 2075, 2080, 2085, 2090, 2095, and 2100, were utilized for the simulations.

The accuracy of historical land-use simulations must be validated before applying the FLUS model to predict future land-use changes. This involves simulating land use for 2015 (with 2010 as the base period) and 2020 (with 2015 as the base period), with accuracy assessed using Kappa and FoM metrics. A Kappa value above 0.75 indicates high reliability, while FoM values typically remain below 0.3 [54].

2.3.3. Trend Test

The Theil–Sen median slope estimation [55,56] and the Mann–Kendall trend test [57,58] were employed to analyze precipitation (P), potential evapotranspiration (PET), and WY trends based on annual time series. The Theil–Sen slope estimation is robust against outliers, while the Mann–Kendall test is distribution-free and insensitive to anomalies. Combining these methods improves efficiency and ensures more accurate trend analysis while effectively accounting for anomalies (significance level = 0.05).

2.3.4. Sensitivity Analysis

From Equations (1) and (2), it is evident that variations in P and PET—both key climatic variables—affect changes in WY. The associated sensitivity coefficients represent the elasticity of WY, defined as the percentage change in WY resulting from a 1% change in either P or PET. These coefficients are instrumental in diagnosing regional hydrological vulnerability. The sensitivity of WY to P and PET can be quantified using the following formulation proposed by [59]:

$$f_P^{\prime }=P{\left(x\right)}^{\left(w-1\right)}{\left(PET{\left(x\right)}^w+P{\left(x\right)}^w\right)}^{\left(1/w\right)-1}$$

(6)

$$f_{PET}^{\prime }=PET{\left(x\right)}^{\left(w-1\right)}{\left(PET\left(x\right)+P\left(x\right)\right)}^{\left(1/w\right)-1}-1$$

(7)

2.3.5. Attributional Analysis

Scenario analysis is the most commonly used method for quantitatively assessing the effects of climate change and land-use change [46]. Three attributional scenarios (i.e., S1, S2, and S3) are established to separate the contributions of climate and land-use change. S1 holds climate and land use constant. S2 involves maintaining a constant climate while altering land use. S3 is the scenario in which both the climate and the land use are changed. The study period is divided into five-year intervals to form several sub-periods. Equations (10) and (11) are used to quantify the impacts of climate and land-use changes on WY in different sub-periods. The specific equations are as follows.

$$∆{WY}_L={WY}_1-{WY}_2$$

(8)

$$∆WY={WY}_1-{WY}_3$$

(9)

$${CP}_L=\left(∆{WY}_L/∆WY\right)\times 100\%$$

(10)

$${CP}_C=1-{CP}_L$$

(11)

where ${WY}_1$, ${WY}_2$, and ${WY}_3$ are the WY for scenarios 1, 2, and 3, respectively; ${∆WY}_L$ is the change in WY due to land-use change; $∆WY$ is the total change in WY; ${CP}_L$ and ${CP}_C$ denote the percentage contribution of climate and land use to the change in WY, respectively.

2.3.6. Geodetector

Geodetector v1.0 is a statistical analysis method based on the principle of spatial stratification heterogeneity, which is mainly used to analyze the spatial differentiation characteristics of geographical phenomena and their driving mechanisms [60]. By quantifying the spatial stratification variance characteristics, the method can effectively identify the key driving factors affecting the distribution of geographical phenomena and determine the explanatory power of each factor. Compared with traditional methods, Geodetector has the advantages of not requiring data to obey specific distribution assumptions and simultaneously evaluating the independent and synergistic effects of multiple drivers. These advantages make it an important tool for spatial heterogeneity attribution analysis.

3. Results

3.1. Performance of the InVEST Model

The Z parameter, an empirical constant in the InVEST model representing local hydrogeological features and precipitation patterns, typically ranges between 1 and 30 and varies by location. This study calculated and compared multi-year mean WY values for different Z parameters against naturalized runoff data. The results indicate that when Z = 13, the long-term average WY is 156 mm, which closely aligns with the observed naturalized runoff of 155 mm. Furthermore, the coefficient of determination (R²) between simulated interannual WY and naturalized runoff reaches 0.71, suggesting good interannual agreement under this parameter setting (Figure 3b). Notably, the 95% confidence interval (CI) of the multi-year average WY at Z = 13 ranges from 129 mm to 180 mm, indicating a relatively narrow uncertainty range and supporting the suitability of Z = 13 for simulating WY in the UYRB.

Figure 4a,b demonstrate that the actual evapotranspiration (ET) derived from the GLEAM dataset and the InVEST model (Z = 13) exhibit similar spatial distribution patterns across the study area. The multi-year average ET values for the period 1990–2020 were 362 mm and 381 mm, respectively. Furthermore, as shown in Figure 4c,d, both datasets indicate an overall increasing trend in ET over time, with regions exhibiting both significant and non-significant upward trends.

In conclusion, the InVEST model produced satisfactory simulation results when the Z parameter was set to 13, indicating that the model configuration is suitable for further application in this study.

3.2. Hydrological Variation and Land-Use Change

3.2.1. Distribution and Trends in Water Yield and Climate Variables

From 1990 to 2020, the UYRB exhibited a statistically significant upward trend in WY (p < 0.05), with an average increase of approximately 3 mm per year and a multi-year mean WY of 156 mm (CI: 129–180 mm). Spatially, WY was generally higher in the southern sub-basins (Figure 5a,b and Figure 6a). Specifically, sub-basins I-1 and I-2 recorded the highest mean WY values of 196 mm and 142 mm, with corresponding 95% CIs of 171–222 mm and 122–164 mm, respectively. In contrast, the northern and central sub-basins exhibited lower WY levels: I-3 averaged 79 mm (CI: 65–94 mm), I-4 averaged 145 mm (CI: 123–169 mm), I-5 averaged 75 mm (CI: 59–91 mm), and I-6 averaged 114 mm (CI: 93–135 mm). All six sub-basins showed significant increasing trends during this period (Figure 6d–i). The southern sub-basins (I-1 and I-2) experienced faster rates of change (3 mm/yr), while the northern and central sub-basins (I-3, I-4, I-5, and I-6) had slower rates (2 mm/yr, 3 mm/yr, 2 mm/yr, and 3 mm/yr, respectively). Figure 7a–d illustrate the structural changes in WY, classified into five categories using the natural breakpoint method. Notably, the proportion of areas in Class V increased significantly from 4.6% to 32.8%, while Classes II, III, and IV exhibited higher dynamic variability.

From 1990 to 2020, the UYRB experienced an average annual precipitation of 538 mm and an average annual potential evapotranspiration of 513 mm (Figure 6b,c). During this period, precipitation exhibited a significantly increasing trend at a rate of 4 mm per year, whereas potential evapotranspiration remained relatively stable. Precipitation was higher in the southern region (Figure 5b), with I-1 and I-2 recording 566 mm and 525 mm, respectively. In contrast, the northern and central sub-basins (I-3, I-4, I-5, and I-6) received 476 mm, 595 mm, 479 mm, and 474 mm, respectively. All sub-basins exhibited increasing precipitation trends, with I-2 showing the highest rate of change (4 mm/yr) and I-3 the lowest (3 mm/yr). Potential evapotranspiration averaged 594 mm and 596 mm in I-3 and I-4, respectively, and 566 mm, 525 mm, 479 mm, and 474 mm in I-1, I-2, I-5, and I-6, respectively. Most sub-basins showed a non-significant upward trend, except for I-5 and I-6, which exhibited a non-significant downward trend (Figure 6d–i).

3.2.2. Spatial–Temporal Changes in Land Use

The UYRB is predominantly covered by grassland and forestland (Figure 8a,b), comprising 86.0% and 6.0% of the total area, respectively (87.3% and 5.7% in 1990; 86.3% and 6.3% in 2020). From 1990 to 2020, land-use changes included a decline in grassland and an increase in forestland and urban areas, with variations observed across sub-basins. Grassland area decreased from 11,223 to 11,009 km², while forestland expanded significantly from 12,322 to 13,522 km², primarily through conversion from grassland (Figure 8a–d). Cropland decreased slightly, with a partial transition to grassland (Figure 8c,d).

Based on the overall trend in land-use change across the study region, the patterns of land use transformation within each sub-basin exhibit notable spatial heterogeneity (Figure 9a–f). In I-1, grassland dominates the landscape, accounting for 97.4% of the area, with primary conversions to forest land (42.4%) and unused land (57.1%). A substantial proportion of cropland (62.5%) was converted to grassland, indicating a clear trend of cropland abandonment and ecological restoration. In I-2, grassland was primarily transformed into forest land (52.7%) and urban land (21.6%), while only 0.8% of forest land experienced conversion, suggesting a coexistence of afforestation and urban expansion. I-3 exhibited the most intensive ecological restoration, with 89.1% of the grassland converted to forest land. In contrast, I-4 showed pronounced urbanization, with 83.2% of newly developed construction land originating from former cropland and grassland. In I-5, 38.1% of cropland and 51.9% of unused land were transformed into grassland and forest land, respectively, reflecting an emphasis on ecological improvement. In I-6, although grassland was largely retained (97.1%), 66.4% of newly added construction land originated from unused land, indicating a distinct trajectory of urban expansion. Overall, I-1 and I-3 are characterized by land restoration efforts, particularly the conversion of farmland to forest and grassland. I-2 and I-4 exhibit varying degrees of urbanization, while I-5 demonstrates a mixed pattern dominated by ecological restoration. Although I-6 experienced a limited land-use change, the direction of urban expansion is evident. These results highlight substantial spatial heterogeneity and divergence in land-use transition pathways across sub-basins.

3.3. Contribution of Climate Change and Land-Use Change to Water Yield Variation

3.3.1. Impacts of Climate Change

As shown in Figure 5 and Figure 6, P and WY in the UYRB share similar spatial patterns and trends, with a strong interannual correlation of 0.97, while PET shows a negative correlation of −0.57 with WY. Across sub-basins I-1 to I-6, P correlates with WY at 0.97 to 0.98, while PET correlations range from −0.50 to −0.68. Using the sensitivity analysis method (Equations (3) and (4)), the mean sensitivity of WY to P is 0.83 (0.70–0.95), with lower values in zones I-4 and I-5 (Figure 10a). The mean sensitivity of WY to PET is generally low at −0.10 (−0.27 to −0.04), with higher sensitivity observed in zone I-1 (Figure 10b).

3.3.2. Impacts of Land-Use Changes

Table 2. Changes in WY per pixel due to land-use type shifts (mm per pixel).

1990/2020	Cropland	Forestland	Grassland	Unused Land	Urban	Sum
Cropland	4	−49	1	27	128	111
Forestland	41	11	77	244	0	373
Grassland	6	−47	15	84	161	219
Unused land	−14	−199	−50	10	108	−145
Urban	−136	0	−147	−108	10	−382
Sum	−99	−283	−104	257	407	177

presents the average annual impact of land-use change on pixel-level WY from 1990 to 2020. Overall, land-use change led to an increase in WY of 177 mm per pixel during this period. Notably, conversions from forest to grassland (+76.89 mm/pixel) and to unused land (+243.64 mm/pixel) substantially increased WY, primarily due to reduced infiltration and enhanced surface runoff following vegetation loss. Urbanization also significantly elevated WY, with transitions from cropland to urban land and from grassland to urban land resulting in increases of +127.94 mm/pixel and +160.96 mm/pixel, respectively. In contrast, the conversion of urban land back to vegetated land types, such as from urban to cropland (−135.98 mm/pixel), led to a marked decrease in WY, underscoring the hydrological benefits of ecological restoration. Transitions between cropland and grassland exhibited relatively minor impacts on WY.

Table 3. Changes in AET per pixel due to land-use type shifts (mm per pixel).

1990/2020	Cropland	Forestland	Grassland	Unused Land	Urban	Sum
Cropland	5.17	38.10	1.20	−18.58	−139.46	−113.57
Forestland	−51.83	3.78	−64.69	−242.31	0	−355.05
Grassland	0.48	61.11	1.44	−64.59	−164.46	−166.03
Unused land	33.23	224.08	70.48	−0.19	−109.27	218.32
Urban	124.12	0	118.76	108.24	−0.88	350.23
Sum	111.17	327.06	127.19	−217.43	−414.07	−66.10

further analyzes changes in actual evapotranspiration (AET) associated with land-use transitions during the same period. Matrix analysis reveals that vegetation restoration has significantly increased AET, with conversions from cropland to forest and cropland to grassland resulting in increases of 38.10 mm/pixel and 1.20 mm/pixel, respectively. Conversely, deforestation—particularly the transition from forest to unused land (−242.31 mm/pixel) and from forest to grassland (−64.69 mm/pixel)—led to substantial reductions in AET. Urban expansion was also associated with a pronounced decrease in AET, with changes from cropland to urban land and grassland to urban land resulting in declines of −139.46 mm/pixel and −164.46 mm/pixel, respectively.

3.3.3. Contributions of Climate Change and Land-Use Change to Water Yield Change

Figure 11a demonstrates that climate change is the dominant factor influencing WY in the UYRB, accounting for an average of 94% over multiple periods, significantly surpassing the contribution of land-use change (6%). This indicates that changes in climatic factors, such as P, have a far greater impact on regional hydrological processes than land-use changes. During the periods 1990–1995, 2000–2005, and 2015–2020, land-use change had a positive but relatively small impact on WY (2%–13%). In contrast, during the periods 1995–2000, 2005–2010, and 2010–2015, land-use change had a negative influence on WY, particularly in 1995–2000, when its contribution was −11%. These results highlight the inconsistent effects of land-use change on WY over different periods. At the sub-basin level (Figure 11b), climate change significantly outweighed land-use change in its contribution across all sub-basins, with climate change contributions ranging from 70% to 122%, further confirming its decisive role in driving WY variations. Sub-basins I-1 and I-2 experienced negative contributions from land-use change (−11% and −22%, respectively), while sub-basins I-3, I-4, I-5, and I-6 saw positive impacts (6%–30%).

Beyond the temporal drivers, we applied the Geodetector model to investigate the spatial heterogeneity of water yield in the UYRB, focusing on the contributions of climate, topography (elevation and slope), and land-use factors. Results show that precipitation is the dominant driver of spatial variation in WY in the upper Yellow River Basin (q = 0.80), followed by cropland and unused land. Elevation and slope also contribute moderately. Interaction analysis reveals that combinations of precipitation with land-use types significantly enhance explanatory power (q = 0.97). Sub-basin comparisons confirm that WY is jointly influenced by climate, land use, and topography.

Factor detection results (Figure 12a) revealed that precipitation (P) was the dominant driver (q = 0.80, p < 0.001), followed by cropland (q = 0.52) and unused land (q = 0.48). Topographic variables such as elevation (DEM) and slope also exhibited moderate explanatory power (q = 0.46), highlighting the importance of terrain in modulating water redistribution.

Interaction detection (Figure 12b) further demonstrated that most combinations of factors produced greater explanatory power than individual factors alone, except for DEM and slope, whose interaction weakened their explanatory capacity. Notably, interactions between precipitation and cropland (q = 0.97), grassland (q = 0.97), and unused land (q = 0.88) significantly enhanced WY explanation. These findings underscore the importance of land use in modulating the hydrological response to climate drivers by regulating infiltration capacity, canopy interception, and surface runoff.

The stark contrast between sub-basins I-1 and I-5 illustrates the spatial heterogeneity of WY in the UYRB. I-1, with the highest WY (196 mm), is characterized by high precipitation (~566 mm), dominant grassland (~97.6%), and steep slopes. These conditions promote orographic precipitation and rapid runoff, limiting soil moisture retention and ET losses. In contrast, I-5 has the lowest WY (75 mm), driven by low precipitation (~479 mm), extensive cropland (~20%), and gentle topography. These factors favor higher soil evaporation and anthropogenic water use, leading to reduced WY. This comparison highlights that WY is not solely controlled by precipitation but by the combined influence of land use and topography, which modulate water partitioning and loss.

3.4. Future Projections of Water Yield

3.4.1. Future Land-Use Scenarios

The FLUS model was utilized to simulate the land use for 2015 and 2020. The simulated and actual land-use data were input into the precision validation module of the FLUS v1.4 software to calculate Kappa and FoM. This resulted in Kappa values of 0.82 and 0.84 for the model identification and validation periods, respectively, and FoM values of 0.21 and 0.23, respectively. These results indicate that the FLUS model’s predictions are reliable and suitable for future land-use simulations.

Under the three future scenarios, land-use trends in the UYRB exhibit significant variations. As shown in Figure 13, grassland remains the dominant land-use type, but its area declines across all scenarios, most notably under SSP126 at a rate of −1194 km²/yr. Forestland area increases or stabilizes, with growth under SSP126 and SSP370 (particularly under SSP126 at 405 km²/yr) and remaining nearly stable under SSP585. Cropland area declines under SSP126 and SSP370, with a sharper decrease under SSP126, but remains stable under SSP585. The unused land area shows a stable or increasing trend, growing under SSP126 and SSP370 (with faster growth under SSP126) and stabilizing under SSP585. The urban area initially increases before declining in all scenarios, with a turning point occurring in 2050.

From the perspective of dominant change types, forestland and grassland are the primary change types, showing a seesaw relationship (Figure 14). Cropland emerges as a significant change type in the later stages under SSP126, while unused land exhibits notable variations across scenarios. Regarding the area of change (Figure 14), SSP126 shows distinct phased changes, with 2065 as a critical turning point. SSP370 demonstrates relatively stable changes, while SSP585 features moderate changes with significant fluctuations in land-use type proportions.

3.4.2. Projected Changes in Water Yield

As shown in Figure 15a,c,d, under future scenarios, both WY and P in the UYRB exhibit significant increases compared to the historical period. At the same time, changes in PET are relatively minor. Under the SSP126, SSP370, and SSP585 scenarios, the average WY is approximately 217 mm, 206 mm, and 201 mm, respectively. The proportion of areas classified as Grade V differs notably, accounting for 40%, 33%, and 33%, respectively (Figure 16). Overall, WY in the UYRB increased relative to the historical period, with SSP 126 > SSP 370 > SSP 585, although the annual water yield varied considerably across scenarios (Figure 15 and Figure 16). At the sub-basin level, WY increases in all sub-basins except for I-5 and I-6. This abnormal phenomenon is mainly attributed to the reduction in regional precipitation. For example, under the SSP126 scenario, the annual average precipitation in the I-5 region decreased from 479 mm in the historical period to 460 mm in the future, while in the I-6 region, it decreased from 474 mm to 426 mm. The reduction in precipitation directly limits the growth of WY (Figure 15b and Figure 16).

3.4.3. Attribution of Future Water Yield Changes

As shown in Figure 17, under the three future scenarios (SSP126, SSP370, and SSP585) for the period 2025–2100, climate change emerges as the dominant factor influencing WY in the basin and its sub-basins. The average contribution of climate change to WY variations is 90.4%, 93.3%, and 97.4% under the SSP126, SSP370, and SSP585 scenarios, respectively, with all scenarios showing a contribution exceeding 90%. Notably, the contribution of climate change increases progressively from SSP126 to SSP585, indicating that the impact of climate change on WY intensifies under more extreme climate conditions. In contrast, the contribution of land-use change to WY demonstrates a declining trend across the scenarios, with average values of 9.1% (SSP126), 5.7% (SSP370), and 3.1% (SSP585) during the 2025–2100 period. Furthermore, the influence of land-use change exhibits significant spatiotemporal variability. Under SSP126, its contribution fluctuates substantially and is relatively unstable. Under SSP370, the variation is more moderate and stable. Under SSP585, the contribution of land-use change shows a pattern of being lower in the earlier period and higher in the later period.

4. Discussion

4.1. Dominant Role of Climate Change in Water Yield Variations

Studies have shown that climate change is critical in driving WY variations [61], particularly through changes in precipitation patterns [62,63]. Using scenario attribution methods, we found that the contribution of climate change to WY in the UYRB exceeded 90% during both historical and future scenarios, indicating that climate change profoundly affects regional hydrological cycles by altering precipitation patterns and evapotranspiration processes. Further analysis of the relationship between P, PET, and WY revealed a significant upward trend from 1990 to 2020, with a strong positive correlation between P and WY (0.97). This suggests that P is the primary driver of WY variation. In contrast, PET showed a negative correlation with WY, indicating that under a warming climate, increased PET may partially offset the positive impact of P on WY, consistent with previous studies [64]. Sensitivity analysis further confirmed that P plays a dominant role, with a sensitivity coefficient of 0.83, while PET had a minimal negative impact (−0.10). These results align with other studies emphasizing precipitation as the primary driver of WY [8,65,66,67].

Under the future scenarios (SSP126, SSP370, and SSP585), increased P led to rising WY across all scenarios, with the largest increase observed under SSP126 (217 mm), suggesting that low-carbon development pathways may enhance water-resource availability. In contrast, under the high-emission SSP585 scenario, although P increased, higher evapotranspiration likely moderated the growth in WY. Furthermore, future scenario analysis revealed that as climate change intensifies (e.g., under SSP585), the contribution of climate factors to WY continues to increase, highlighting the amplified influence of climatic drivers under extreme conditions.

4.2. The Secondary Role of Land-Use Change in Water Yield

WY variations are not only driven by climatic factors but are also influenced by land use [3,16]. Land use affects WY by altering land-cover types and hydrological processes, such as infiltration, evapotranspiration, and soil water storage [68]. However, the magnitude and direction of these impacts are jointly constrained by regional natural conditions, land-use conversion types, and temporal scales. From 1990 to 2020, the primary land-use changes in the UYRB included forest expansion, grassland reduction, accelerated urbanization, and decreased unused land. Overall, forest expansion reduced WY by increasing AET, while the conversion of unused land and grassland into cropland or urban land generally enhanced WY by reducing AET or increasing runoff.

Quantitative analysis from Table 2 confirms that forest expansion substantially elevated AET, especially through transitions such as cropland to forest (+38.10 mm/pixel) and grassland to forest (+61.11 mm/pixel). These increases in AET led to reductions in WY due to enhanced canopy interception and transpiration by deep-rooted forest species [3,15,67,69,70]. Forest expansion reduces WY by enhancing evapotranspiration and precipitation interception [67]. This finding contrasts with the traditional belief that forests contribute to water conservation, indicating that in semi-arid regions, forest expansion may reduce water availability rather than increase it [71]. Conversely, deforestation caused dramatic reductions in AET—for instance, forest-to-unused land and forest-to-grassland—which likely resulted in increased WY. Similarly, urban expansion led to substantial declines in AET, thereby increasing WY due to reduced infiltration and elevated surface runoff [72,73]. However, this hydrologic gain may be counteracted over time by increasing water consumption for urban use. Transformations involving unused land were particularly impactful: conversion to forest reduced WY, while conversion to cropland or grassland increased WY, albeit with a moderate rise in AET. These findings suggest that while vegetation restoration promotes ecological benefits, it can also reduce water yield in water-limited basins, underscoring the trade-offs between ecological and hydrological objectives.

From a temporal perspective, the contribution of land use to WY varied significantly across different periods. During the periods 1990–1995, 2000–2005, and 2015–2020, land use had a positive contribution to WY (13%, 6%, and 2%, respectively), mainly because the positive effects of unused land conversion and urban expansion partially offset the negative impacts of forest expansion. However, in 1995–2000 and 2005–2010, land use negatively impacted WY (−11% and −1%), primarily due to the dominant hydrological effects of forest expansion and urban contraction.

Future scenario projections indicate that as climate change intensifies, the contribution of land use to WY is likely to decrease further [70]. Under the SSP126, SSP370, and SSP585 scenarios, the contribution of land use to WY is projected to be 9.1%, 5.7%, and 3.1%, respectively, showing a declining trend as climate change intensifies.

In conclusion, while land use has a relatively minor impact on WY in the UYRB compared to climate change, its influence in specific regions and periods remains significant. To ensure sustainable water-resource utilization, future water-resource management should consider the combined effects of climate change and land use, particularly in ecological conservation, land-use planning, and water-resource regulation.

4.3. Limitations of the Study

Although our study provides significant insights into the spatiotemporal dynamics of water yield (WY) under climate and land-use change scenarios in the upper Yellow River Basin (UYRB), several inherent limitations must be acknowledged, primarily related to the chosen modeling approach and assumptions.

Firstly, the InVEST model simplifies hydrological processes through the application of the Budyko framework and annual water balance calculations. The Budyko framework inherently integrates water–energy interactions, capturing the essential coupling between available water (precipitation) and energy (potential evapotranspiration) to determine actual evapotranspiration and water yield. This physically based underpinning provides robustness to the model despite its simplicity. The InVEST model assumes that all water yield directly reaches watershed outlets without differentiating between surface runoff, baseflow, and groundwater flow components [6,23]. While this simplification could mask critical hydrological responses, particularly distinguishing baseflow from surface runoff, vital for detailed water management and flood mitigation strategies, the primary goal of our study was to analyze total basin-scale water yield (WY = P − ET). In calibrating the model, we utilized multi-year average runoff data at the basin outlet. At multi-year scales, the change in watershed storage is negligible, and the water balance equation simplifies to runoff = precipitation − evapotranspiration. Therefore, it is reasonable and justified to calibrate the InVEST model using multi-year average runoff data. Nonetheless, future research could adopt integrated modeling frameworks coupling InVEST with physically based models (e.g., SWAT or VIC), explicitly considering baseflow processes and groundwater–surface water interactions, as demonstrated effectively in recent studies [6,28].

Secondly, the evapotranspiration (ET) calculations in InVEST use the Hargreaves equation, a relatively simple empirical method, rather than the more physically detailed Penman–Monteith (PM) equation. However, the PM equation requires a large amount of meteorological data, which may be difficult to obtain or impractical in regions with sparse meteorological stations (such as high-altitude, cold regions, like the UYRB). In contrast, the Hargreaves method, which requires fewer input data, has significant practical advantages in large-scale scenario analysis and data-scarce environments. Although we validated the evapotranspiration results using remote sensing GLEAM evapotranspiration data (corrected by flux tower observations), future studies should directly compare the Hargreaves equation with the Penman–Monteith equation under long-term climate scenarios to better quantify uncertainties and enhance model reliability.

Thirdly, despite the availability of relatively abundant datasets in the UYRB region that could theoretically support physically based hydrological models, our rationale for selecting the InVEST model lies in its widespread use and demonstrated suitability for effectively distinguishing the impacts of climate change and land-use change, as extensively validated in previous studies [6,28]. Moreover, the raster-based evapotranspiration (ET) and water yield (WY) outputs generated by the InVEST model facilitate detailed grid-level analysis and enable direct spatial correlations with land-use changes, thereby enhancing the depth and relevance of large-scale studies.

Lastly, the FLUS model assumes land-use transition rules remain constant over time, potentially limiting its effectiveness in capturing long-term shifts driven by evolving socioeconomic factors and policy interventions [52,54,74]. Future research should integrate dynamic socioeconomic scenarios into land-use simulations to improve predictive robustness and policy relevance.

In summary, although our selected models provide robust and practical insights for evaluating long-term WY dynamics under integrated climate–land scenarios, acknowledging these inherent simplifications and assumptions is essential. Future studies utilizing comprehensive modeling approaches, physically based hydrological models, and detailed ET calculation methods will significantly enhance the robustness and accuracy of hydrological assessments and ecosystem service evaluations.

4.4. Implications and Suggestions

Studies have shown that climate change is the dominant factor affecting water yield in the UYRB, while land-use change also plays a significant role in specific regions and periods. Under the backdrop of ongoing global warming, both regional climate conditions and land surface characteristics in the UYRB are expected to undergo substantial changes in the coming decades and by the end of the 21st century, leading to corresponding and pronounced variations in regional water yield. Therefore, based on the current conditions and future development trends of the UYRB, we propose the following recommendations for future water-resource management:

First, strengthening the monitoring of climatic factors, especially precipitation, at both regional and sub-basin scales. Long-term and continuous monitoring of climate variables forms the scientific basis for assessing future climate change. Most areas of the UYRB are located in high-altitude, cold regions, and our findings, consistent with previous studies, indicate that the region is highly sensitive to climate change, with precipitation being the primary driver of water yield variation. Considering the still sparse distribution of meteorological stations in the basin, it is imperative to optimize and install additional automatic weather stations.

Second, enhancing the quantitative prediction of future climate change in the region. As future climate trends will largely determine future water yield, it is essential to utilize advanced technologies such as Earth system models and deep learning to conduct quantitative projections of key climate variables and extreme events. Reducing prediction uncertainty is fundamental for developing effective adaptation strategies.

Third, improving the monitoring, regulation, and planning of land-use processes. Integrated approaches, including field surveys, satellite remote sensing, and unmanned aerial vehicles, should be employed to achieve full-process monitoring of major land cover types. Following the principle of ecological priority, it is necessary to regulate the scale and pace of cropland and urban land expansion appropriately. Optimizing land-use planning, enhancing ecosystem management, and promoting region-specific ecological restoration are crucial. Additionally, it is crucial to prevent excessive and disordered forest expansion, which could lead to the overconsumption of water resources.

Fourth, promoting low-carbon and green development and strengthening climate resilience strategies. The projected water yield under different development scenarios varies significantly; low-emission pathways (e.g., SSP126) are associated with higher water yield compared to high-emission scenarios (e.g., SSP585), thereby mitigating the negative impacts of water scarcity. Thus, efforts should be made to advance clean energy transitions, low-carbon industrial transformation, and the enhancement of ecological carbon sinks to reduce anthropogenic disturbances to the climate system. Meanwhile, in order to address potential extreme hydrological events (e.g., droughts and floods) triggered by climate change, an adaptive water resource management system must be established. This includes improving infrastructure resilience, refining risk early warning systems, and ensuring a stable water supply in the face of climate variability.

Fifth, to adhere to the principle of “development within water limits” and formulate a development model oriented toward water resource security. The water resources of the upper Yellow River Basin are crucial for the ecological security and socioeconomic development of downstream regions. However, future uncertainties concerning water yield will directly influence the sustainability boundaries of the basin. Therefore, water-resource carrying capacity should serve as a core constraint in regional development planning, ensuring the coordinated development of water resources, the economy, and ecosystems. At the same time, a rigid constraint mechanism based on “water-determined urbanization, industrialization, and population” should be established. Water-resource red lines must be integrated into government performance evaluations and regional development assessments to shift from a “resource-driven” to a “resource-constrained” development paradigm, thereby improving the region’s capacity to adapt to climatic and hydrological uncertainties.

5. Conclusions

This study conducted a comprehensive quantitative assessment of the historical and future variations in water yield (WY) in the ecologically sensitive upper Yellow River Basin (UYRB) from 1990 to 2100, under the combined effects of climate change and land use transformation. The main conclusions are as follows:

(1)

Historical Trends (1990–2020)

Water yield in the UYRB exhibited a statistically significant increasing trend, with an average annual growth rate of 3 mm/yr and a basin-wide mean WY of 156 mm. Spatial heterogeneity was evident, with higher WY observed in the southern sub-basins and lower values in the northern parts. All six sub-basins demonstrated a consistent upward trend in WY, suggesting a basin-wide hydrological response to changing climatic conditions.

(2)

Dominant Role of Climate Change

Attribution analysis revealed that climate change has been the primary driver of WY dynamics during the historical period, contributing 94% of the total change, whereas land-use change accounted for only 6%. Precipitation was the major positive driver, while increasing potential evapotranspiration had a moderate but negative impact. The impact of land-use change exhibited strong spatiotemporal heterogeneity, depending on land cover transitions such as forest expansion, urbanization, and cropland transformation. These processes differentially influenced WY due to their contrasting effects on evapotranspiration and infiltration.

(3)

Future Projections (2025–2100):

Under all three SSP scenarios (SSP126, SSP370, and SSP585), WY is projected to increase, reaching 217 mm, 206 mm, and 201 mm, respectively. However, the extent of this increase varies with emission intensity. Low-carbon development under SSP126 may enhance WY by stabilizing evapotranspiration rates, whereas high-emission scenarios (e.g., SSP585) could limit WY gains due to intensified atmospheric water demand. Concurrently, the relative contribution of land-use change to future WY variation is expected to decline further to 9.1%, 5.7%, and 3.1% under SSP126, SSP370, and SSP585, respectively.