Four main datasets were employed in the improved CCW model: vegetation data (NDVI), climate data (precipitation, temperature, shortwave radiation, vapor pressure deficit, and atmospheric pressure), land use and land cover (LULC), and [CO₂]. The monthly NDVI dataset used in this study (Table 1) was derived from a daily 0.05° gap-free NDVI dataset in China (10.6084/m9.figshare.c.7002225.v1, Li et al., 2024a), which was developed from the NOAA's daily NDVI dataset, applying effective data recognition and spatiotemporal gap-filling techniques. The dataset spans 1981–2023 and provides a spatial resolution of 0.05°, and we used bilinear interpolation to generate the dataset with a spatial resolution of 0.1°.

Climate data (Table 1), including precipitation, air temperature, surface downward shortwave radiation, relative humidity, and atmospheric pressure, were sourced from the China Meteorological Forcing Dataset (CMFD) at the National Tibetan Plateau Data Center (TPDC) of the Institute of Tibetan Plateau Research, Chinese Academy of Sciences (He et al., 2020). The dataset spans 1979–2018 and provides a spatial resolution of 0.1° and temporal resolutions at 3 h, daily, monthly, and annual scales. As the dataset did not provide vapor pressure deficit (VPD), we calculated VPD using the method from Howell and Dusek (1995), based on atmospheric pressure, temperature, and relative humidity.

LULC data (Table 1) were obtained from the Zhang et al. (2024) global dataset, which provides consistent multi-temporal global LULC maps at 30 m spatial resolution for 1985–2022. The dataset includes 35 fine-resolution LULC types. For the purposes of this study, and to facilitate LULC change analysis, we merged these 35 LULC types into 17 types using the IGBP classification, based on the method by Yang et al. (2017). Four primary LULC types – cropland, forest, grassland, and bare land – were determined following the method described by Mu et al. (2013). The data were resampled to the 0.1° spatial resolution, ensuring compatibility for modeling within the modified CCW framework.

[CO₂] data were sourced from the Mauna Loa Observatory (MLO), Hawaii (20° N, 156° W) (https://gml.noaa.gov/ccgg/trends/data.html, last access: 1 June 2026), with yearly observations used to represent national [CO₂] levels due to the minimal spatial variation in [CO₂] across China (Table 1). These datasets were then used to drive the improved CCW model.

In this study, the hydrological data for model validation from 145 hydrological stations (Fig. 1), each with at least 15 years of continuous data since 1982, was collected from the Hydrological Bureau of the Ministry of Water Resources of China (https://www.mwr.gov.cn/english/, last access: 18 June 2026). Annual runoff data were calculated from the daily runoff and the catchment area controlled by each hydrological station.

The Coupled Carbon and Water (CCW) model is a remote-sensing driven ecosystem model that mechanistically links terrestrial carbon uptake and water dynamics (Zhang et al., 2016b), and its detailed information is provided in Appendix A. In this study, we improve the CCW model by replacing the static underlying water use efficiency (UWUE) in the original CCW model with a dynamic ecosystem-scale WUE that responds to changing [CO₂], vapor pressure deficit (VPD), atmospheric pressure, and vegetation structure.

In the improved CCW model, evapotranspiration (ET) is calculated as the ratio of GPP to WUE: 1ET=GPPWUE where WUE is defined as the ratio of GPP (Gross primary productivity) to total ecosystem evapotranspiration. GPP is estimated as: 2GPP=APAR×ε=PAR×FPAR×εpot×Rs×Ts×Ws where APAR is the absorbed photosynthetically active radiation (MJ m⁻²), which is calculated as the product of incident photosynthetically active radiation (PAR) and the fraction of PAR absorbed by vegetation (FPAR). PAR is typically assumed to be 45 % of the total shortwave radiation (Running et al., 2000), and FPAR is determined by the normalized difference vegetation index (NDVI) (Sims et al., 2005). ε is the realized light-use efficiency (g C MJ⁻¹), which is calculated by multiplying the potential light-use efficiency (εpot) and environmental scalars for diffuse radiation (Rs), temperature (Ts), and moisture stress (Ws).

WUE is estimated using the WEC (Water Efficiency and Carbon) equation proposed by Cheng et al. (2017): 3WUE=Ca×Pa1.6VPD+g1VPD1-exp⁡-k⋅LAI1-fi where Ca is atmospheric CO₂ concentration (mol(CO₂) mol⁻¹(air)); Pa is atmospheric pressure (kPa); VPD is vapor pressure deficit (kPa); g1 is an empirical parameter of the Ball stomatal conductance model; k is the radiation extinction coefficient, typically set at 0.6; LAI is the leaf area index; and fi is fraction of interception to total evapotranspiration. In this study, the interception evaporation factor (fi) was set to zero, since Cheng et al. (2017) reported that canopy interception account for a minor portion of total evapotranspiration at annual scales. Given that the improved CCW model focused on yearly water yield (WY) dynamics, neglecting interception loss reduces model complexity without substantially affecting WY estimation.

In order to ensure the consistency of NDVI and LAI trends, LAI is calculated using NDVI (Gutman and Ignatov, 1998): 4LAI=-2ln⁡1-fNDVIfNDVI=NDVI-NDVI0NDVI1-NDVI0 where NDVI0=0.04, NDVI1=0.52

Under this framework, ET should be interpreted as an effective annual ecosystem-scale evapotranspiration suitable for long-term WY attribution, rather than a detailed partitioning of individual ET components. Compared with the original CCW model, this modification allows ET to vary with dynamic WUE and therefore enables the model to represent [CO₂]-induced physiological regulation in addition to vegetation structural and climatic effects.

Finally, the water yield (WY) is calculated as the difference between precipitation (P) and ET: 5WY=P-ET

On an annual scale, WY can be approximately equal to runoff, as changes in soil water storage over long periods (one year or longer) are considered negligible (Xiao et al., 2020; Zhang et al., 2021b). Thus, the attribution of WY can also be considered as the attribution of runoff. Accordingly, in this study, WY is used as the modelled output, while the term “runoff” is reserved for observed streamflow or literature values explicitly labelled as such. This approximation is most suitable for annual to multi-year analyses in regions without substantial long-term groundwater depletion or strong reservoir regulation.

In this study, we proposed an attribution analysis framework based on scenarios simulation to explore the combined and individual effects of climate, vegetation, and [CO₂] change on WY. Four scenarios were designed (Table 2). Scenario 1 (Actual) aimed to validate the improved CCW model and estimate the combined effects of climate, vegetation, and [CO₂] change on WY by allowing all factors to vary from 1982 to 2017. Scenario 2 (Vegetation Change) focused on estimating the direct effects of vegetation change on WY by allowing vegetation factors (NDVI and LULC) to vary while keeping climate and [CO₂] fixed at 1982 levels. In this case, the trend in WY obtained reflects the impact of vegetation change alone. Scenario 3 (Climate Change) aimed to estimate the direct effects of climate change on WY by allowing climate factors (precipitation, temperature, relative humidity, solar radiation, and atmospheric pressure) to change, while fixing vegetation and [CO₂] at 1982 levels. This scenario helps isolate the effects of climate change on WY. Scenario 4 ([CO₂] Change) was designed to estimate the direct effects of [CO₂] change on WY by varying [CO₂] levels from 1982 to 2017, while climate and vegetation factors were fixed at 1982 levels. The resulting WY trend reflects the impact of [CO₂] change alone. The resulting WY series under each scenario represents the direct impact of the corresponding driver under the assumption that the other drivers are fixed.

For each scenario, the long-term “trend” in annual WY over 1982–2017 was quantified using the Theil–Sen slope of the annual WY series (Sen, 1968; Theil, 1992). The relative contributions of climate, vegetation, and [CO₂] to changes in WY were calculated using the following formula (Ma et al., 2023; Wang et al., 2022): 6RCvegetation=trendvegetationtrendvegetation+trendclimate+trendCO2×100%RCclimate=trendclimatetrendvegetation+trendclimate+trendCO2×100%RCCO2=trendCO2trendvegetation+trendclimate+trendCO2×100% where trendvegetation, trendclimate, and trendCO2 denote the Theil–Sen slopes of the annual WY series under vegetation, climate, and [CO₂] scenarios, respectively. These slopes represent the long-term rates of WY change attributable to each driver within the scenario framework; the relative contributions (RCvegetation, RCclimate, and RCCO2) are expressed as percentages, indicating the normalized magnitude of the scenario-derived direct effects.

At each grid, the dominant driver of WY change is identified as the factor with the largest absolute relative contribution among climate, vegetation, and [CO₂]. When the absolute difference in relative contributions between two factors do not exceed 5 %, these two factors are considered as joint significant contributors to the changes in WY at that grid (Jia et al., 2022). This criterion is used to distinguish grids dominated by a single factor from those where two drivers contributed comparably to WY change.

Furthermore, elasticity analysis is implemented to quantify the sensitivity of WY to individual drivers. Elasticity is defined as the relative change in WY caused by a 1 % perturbation in a given driver (Yang and Yang, 2011). We specifically focus on precipitation because, despite not always having the highest sensitivity, it is integral to the hydrological cycle and essential for assessing water yield (WY) under various climate change scenarios (Liu et al., 2017). Specifically, for each driver (x), including precipitation, NDVI, and [CO₂], the baseline S1 series is increased by 1 % over the entire 1982–2017 period while all other factors are held unchanged. Then the elasticity coefficient is estimated as: 7ΔRxRx=WYmeanx-WYmeanS1WYmeanS1 where (x) represents precipitation, NDVI, or [CO₂]; WYmeanS1 is the mean annual WY under the baseline scenario S1 during 1982–2017; and WYmeanx is the mean annual WY after increasing driver (x) by 1 % while keeping all other factors unchanged.

Mathematically, the elasticity coefficient is defined as the runoff change rate divided by 1 %, and the formula is as follows: 8εx=ΔRxRxΔxx=ΔRxRx1%

Generally, while the scenario analysis above has identified which factors are most influential based on their relative contributions, the elasticity coefficients allow us to explain why these factors are critical by demonstrating their respective impacts on WY through sensitivity analysis. By separating the sensitivity of WY from the historical magnitude of driver changes, the elasticity analysis extends the scenario-based attribution and provides a basis for approximating future WY responses when projected changes in precipitation, vegetation, or [CO₂] are available.

Figure 2 demonstrates the trends of annual precipitation, air temperature, relative humidity, atmospheric pressure, solar radiation, and NDVI across China during 1982–2017. Annual precipitation change exhibited a clear spatial distribution pattern, specifically decreases in central China, including the middle reaches of the Yellow River and the Yangtze River basins, and increases in the northwest and southeast. Air temperature exhibited a consistent warming trend across China. In contrast, relative humidity generally decreased across most China. Atmospheric pressure remained relatively stable. Regarding solar radiation, decreases were in northern China, while an increase was in southern regions. The decreasing solar radiation in northern China is likely due to increased aerosol concentrations (Liang et al., 2024). NDVI showed a significant increasing trend, which indicates an overall enhancement in vegetation growth across China. This trend was especially prominent in central and eastern regions, including the Yellow River Basin and the Yangtze River Basin. In these regions, LULC changes, such as afforestation and agricultural practices, likely contributed to the observed increases in NDVI (Chen et al., 2019).

Significant changes in land use and land cover (LULC) occurred in China during 1982–2017, as illustrated in Fig. 3. Although the overall percentage distribution of major land cover types, namely grasslands, forests, croplands, and bare lands, remained relatively stable, these four categories dominated the landscape, with most changes concentrated within them. Notably, the transitions among these categories were characterized by mutual conversions, particularly from bare land to grasslands (Fig. 3). Spatially, the changes exhibited distinct regional patterns. In southern China, LULC changes were mainly characterized by the conversion of land to forests and grasslands. In contrast, the northeastern regions exhibited more complex transformations, with some areas shifting to bare land and croplands (Fig. 3).

As shown in Fig. 4a and b, the observed annual water yield (WY) and the simulated annual WY by the improved CCW model showed strong linear correlations (R2 = 0.7), with the regression line slope being 1.45, R2 being 0.7, and RMSE being 9.54 mm yr⁻¹. By contrast, the initial model without WUE showed weaker skill (slope = 1.45, R2 = 0.68, RMSE = 9.62 mm yr⁻¹), indicating that explicitly representing [CO₂]-induced regulation of water-use efficiency measurably improves accuracy and reduces bias.

The estimated annual WY trends had distinct spatial patterns (Fig. 4c), which closely aligned with that of precipitation. Specifically, decrease trends in WY occurred in the central regions of the Yellow River Basin and the middle section of the Yangtze River Basin, while increase trends were found in other regions, with the southeast exhibiting the highest rate of increase.

Figure 5 shows the distribution of WY changes caused by climate, vegetation, and [CO₂] changes, integrating both absolute magnitude (Fig. 5a, c, e) and relative dominance (Fig. 5b, d, f) of their contributions. Climate-driven WY changes exhibited marked spatial heterogeneity, with absolute increases exceeding 15 mm yr⁻¹ in southeastern China (Fig. 5a), corresponding to 60 %–90 % relative contributions (Fig. 5b). Central basins showed contrasting declines of 0–6 mm yr⁻¹ under climate forcing, while northeastern transitional zones displayed mixed positive/negative absolute changes (Fig. 5a) despite maintaining 40 %–70 % relative climate dominance (Fig. 5b). This spatial heterogeneity aligned with precipitation change patterns (Fig. 2a).

Vegetation-mediated WY reductions reached 0–6 mm yr⁻¹ (Fig. 5c), accompanied by 0 %–60 % relative contributions (Fig. 5d). These effects originated from enhanced evapotranspiration through LULC changes and vegetation greening, particularly pronounced in central China. Specific regions in the Yangtze, Yellow, and northeastern rivers showed vegetation-driven relative contributions reaching 40 %–60 % (Fig. 5d). [CO₂] effects generated limited direct absolute impacts (< 5 mm yr⁻¹, Fig. 5e) but exerted 10 %–40 % relative influences (Fig. 5f) through stomatal closure mechanisms. This process partially counteracted vegetation-related WY decreases in transitional climates like northeastern China, where competing drivers created complex ecohydrological interactions (Fig. 5).

Figure 6 illustrated the spatial distribution of WY trend drivers over the past four decades. Climate change was the dominant factor of WY variation in more than 70 % regions, mainly in the Northwest, Southwest, Southeast, Pearl River basins, and other parts of the Yangtze and Yellow River basins. Vegetation changes ranked as the secondary control, dominating WY changes in parts of the Yangtze, Yellow, Songhua, Liao, and Hai Rivers. Remarkably, it was shown that the region where vegetation and [CO₂] had the dominant influence mainly distributes within precipitation ranges of 400–1600 mm. [CO₂]-induced effects were least influential at a national scale. This three-tiered hierarchy – climate changes as the primary forcing, vegetation changes as the secondary control, and [CO₂] effects as a localized modifier – reveals how hydrological regimes govern the spatial succession of dominant drivers across China's diverse ecohydrological gradients.

The elasticity of WY to precipitation (εP), NDVI (εNDVI), and [CO₂] (εCO₂) exhibits distinct spatial patterns in Fig. 7. Nationally averaged elasticity coefficients showed that a 10 % increase in precipitation, [CO₂], and NDVI altered WY by 15.5 % (εP=1.55), 5.5 % (εCO2=0.55), and -4.4 % (εNDVI = -0.44), respectively, indicating that, in terms of the sensitivity of runoff to changes in each factor, the ranking was precipitation > [CO₂] > NDVI. The historical relative changes in precipitation (0.48 % yr⁻¹), [CO₂] (0.49 % yr⁻¹), and NDVI (0.67 % yr⁻¹), suggest that precipitation increased WY by 27.2 %, [CO₂] increased WY by 9.97 %, and NDVI decreased WY by 10.7 % over 1982–2017. Therefore, although [CO₂] had a higher elasticity than NDVI, its historical contribution to WY change was slightly smaller than that of vegetation.

The elasticity coefficients of precipitation (εP), [CO₂] (εCO₂), and vegetation (|εNDVI|) all exhibited a coherent latitudinal decline across China's river basins, showing systematically higher sensitivity in northern regions than southern counterparts. Quantitatively, εP decreased from 2.09 in the Songhua River basin to 1.15 in the Southeastern Basin, accompanied by similar reductions in |εNDVI| (from 0.76 to 0.13) and εCO₂ (from 1.08 to 0.16) (Table 3).

A distinct abrupt transition zone in elasticity coefficients was identified around 33° N, closely aligning with China's traditional North-South physiographic divide. Around the zone, elasticity coefficients exhibited an abrupt decline from the Yellow River Basin to the Yangtze River Basin. Specifically, the Yellow River Basin showed higher sensitivities to precipitation (εP=1.87), [CO₂] (εCO2=0.86), and NDVI (εNDVI = -0.53), which were approximately 1.4, 2.8, and 2.8 times greater, respectively, than those in the Yangtze River Basin (εP=1.31, εCO2=0.31, εNDVI = -0.19).

This study developed an attribution framework based on the improved CCW model to separate the effects of climate change, vegetation structural change, and [CO₂]-induced physiological changes on WY changes. A key strength of this framework is that it explicitly links vegetation dynamics and [CO₂] effects to actual evapotranspiration at grid scale, rather than representing them indirectly through empirical catchment parameters or potential evapotranspiration adjustments.

First, the improved framework addresses a major limitation of Budyko approaches. Traditional Budyko approaches often attribute vegetation effects to temporal variations in the Budyko parameter n by either statistically regressing the Budyko parameter against vegetation proxies such as NDVI (Liu et al., 2024; Tan et al., 2023) or simplistically equating the Budyko parameter to vegetation effects (Li et al., 2020b; Zhou et al., 2023). In addition, some Budyko approaches account for [CO₂] effects by adjusting potential evapotranspiration (Liu et al., 2024). These treatments can conflate vegetation change and [CO₂] change with other catchment properties embedded in n (Zeng et al., 2020). Moreover, when [CO₂] effects are represented through PET adjustments, their influence on WY is transmitted through the Budyko function and further constrained by n, making it difficult to isolate [CO₂]-induced physiological effects from vegetation structural change and other land-surface controls. In contrast, our framework estimates actual ET through a carbon–water coupling structure and explicitly separates the effects of climate, LAI-related vegetation structural change, and [CO₂]-driven physiological changes, thereby providing a more direct basis for attributing WY changes than Budyko approaches.

Second, compared with the original CCW model using a static UWUE, our framework replaces it with a dynamic WUE formulation that accounts for long-term WUE changes associated with variations in [CO₂], VPD, and vegetation structure. Elevated [CO₂] reduces stomatal aperture, thereby lowering stomatal conductance and transpiration flux while only modestly increasing carbon assimilation, leading to an overall enhancement in WUE. This process is represented by the Medlyn-type stomatal conductance model (Medlyn et al., 2011), which links photosynthetic rate (A), transpiration (T), and vapor pressure deficit (D) as: 9AT=CaPa1.6D+g1D where Ca is atmospheric CO₂ concentration, Pa is air pressure, D is vapor pressure deficit, and g1 is an empirical slope parameter that quantifies plant sensitivity to [CO₂] and humidity. This formulation also indicates that WUE varies with [CO₂] and atmospheric water demand, rather than remaining constant over time. Incorporating this relationship into the CCW framework enables the model to distinguish NDVI-driven vegetation change effects on WY from [CO₂]-induced stomatal effects. Therefore, this improvement is not merely a statistical refinement, but also a process-based modification that helps explain why [CO₂] can partly offset vegetation-related WY reductions in some regions.

Third, the grid-scale implementation of the framework allows spatial heterogeneity in attribution results to be identified within basins. Many previous runoff attribution studies have been conducted at the basin scale (Liu et al., 2024, 2017; Yang et al., 2022), which is useful for summarizing overall basin responses but may mask substantial intra-basin differences. By conducting attribution at the grid scale, our framework captures spatial heterogeneity in WY controls within the same basin, distinguishing areas dominated by climatic variability from those more strongly affected by vegetation greening. This spatially explicit attribution provides more detailed and differentiated information for regional water resource assessment and ecological restoration planning within a basin.

Our findings highlight climate change as the dominant driver of water yield (WY) changes (contributing > 70 %), which is consistent with previous assessments (Table 4). Climate impacts dominate in the Northwest and Southwest River Basins, as well as parts of the Yangtze, Yellow, Southeast, and Pearl River Basins. In contrast, vegetation and CO₂ dominate WY change in central China (portions of the Yangtze, Yellow, Songhua, Liao, and Hai River basin). The spatial pattern is slightly distinct from previous studies (Table 4), with noticeable discrepancies in some regions. These discrepancies may arise from differences in study periods, as vegetation effects have become more pronounced after 2000 in several regions but may be weakened when averaged over longer historical periods. This basin-scale contrast is also consistent with a broader hydroclimatic gradient, as vegetation and [CO₂] effects become relatively more important in the intermediate precipitation zone of 400–1600 mm. The relative importance of vegetation and [CO₂] is not randomly distributed but is concentrated along transitional hydroclimatic zones, where water availability is sufficient to support vegetation growth yet remains sensitive to changes in evapotranspiration.

Elasticity analysis (Sect. 3.4) revealed distinct sensitivities of WY to environmental drivers: precipitation exhibited the highest elasticity coefficient for the whole China (εP = 1.55), followed by [CO₂] (εCO2=0.55) and NDVI (εNDVI = -0.44). However, spatial analysis showed that vegetation and [CO₂] collectively dominated WY changes in the 400–1600 mm precipitation zones, despite their lower sensitivity rankings. The contribution of each driver is determined by its elasticity multiplying the magnitude of its change. In the 400–1600 mm precipitation zones, NDVI displayed a larger relative temporal variation compared with precipitation (Fig. 8), which fluctuated within a narrower range. Therefore, high elasticity does not necessarily indicate high historical contribution. The larger relative change in NDVI enhanced the contribution of vegetation to WY change despite its lower elasticity, whereas the historical contribution of [CO₂] was limited by its slower accumulation rate (0.49 % yr⁻¹) despite its relatively high elasticity.

Although the historical contribution of [CO₂] was limited by its slow increase during 1982–2017, future [CO₂] may substantially amplify its effect on WY. Projections from the CMIP6 SSP585 scenario indicate [CO₂] will rise at 2.34 % yr⁻¹, which is nearly five times faster than the historical rate (Cheng et al., 2022). Using the historical elasticity coefficient as a simple sensitivity-based approximation, the increase of [CO₂] would drive a +1.29 % yr⁻¹ increase in annual WY by 2100. By comparison, future NDVI will increase at approximately 0.59 % yr⁻¹ under CMIP6 SSP585 (Li et al., 2025), its effect on WY would be only -0.26 % yr⁻¹. Therefore, hydrological effect of rising [CO₂] may be much larger than the direct effect of future vegetation greening and therefore should not be ignored in future WY assessments. However, this estimate assumes that the historical [CO₂] elasticity remains unchanged and does not account for possible counteracting effects, such as rising VPD, drought stress, or saturation of WUE under elevated [CO₂] (Adams et al., 2020; Li et al., 2023).

From a policy perspective, these spatial contrasts have distinct implications for regional water management. In regions primarily affected by vegetation change, such as the Yangtze and Yellow River basins, priority should be given to optimizing vegetation composition and preventing overgreening that may reduce runoff. Conversely, in regions primarily affected by climate change, such as Northwest and Southeast China, adaptive measures that emphasize precipitation variability, water storage capacity, and drought resilience are crucial. Tailoring water management strategies to these distinct regional conditions can enhance the effectiveness of both ecological restoration and climate adaptation programs across China.

This study used an improved CCW model to attribute annual WY changes in China to climate change, vegetation structural change, and [CO₂]-induced physiological effects, providing a more mechanistic understanding of vegetation and [CO₂] effects of hydrological responses. However, several limitations remain that need to be addressed in future work to improve the accuracy and robustness of the results.

First, the adopted WUE formulation is defined at the ecosystem scale, but the present implementation simplifies the interception term by assuming fi=0. Soil evaporation is still implicitly represented through the LAI-based partitioning term, whereas canopy interception is not explicitly allowed to vary across space or time. This simplification may bias the absolute magnitude of ET and WY, particularly in humid and forested regions, although its influence on long-term WY trends is expected to be small. Nevertheless, given the small interannual variation of fi (Zhao et al., 2022), its influence on runoff trends is negligible in our study (Cheng et al., 2017). Future work should prioritize its explicit calculation.

Second, the attribution framework treats climate, vegetation, and [CO₂] as separate drivers, whereas these factors interact with each other in real ecohydrological systems. Vegetation exhibits interactions and feedbacks with climate, making it difficult to strictly separate the impacts of climate change, vegetation dynamics, and [CO₂] on hydrological responses. Changes in vegetation, such as NDVI, reflect a combination of climate change, human activities (e.g., reforestation and irrigation), and natural vegetation growth. Additionally, vegetation greening in upwind regions can increase atmospheric moisture, potentially enhancing precipitation downwind (Zhang et al., 2021a), which may counteract some of the negative impacts of increased evapotranspiration on local WY. Although the climate data used in our model may implicitly capture some of these feedbacks, they cannot be explicitly separated in this analysis. This limitation also affects the simple elasticity-based SSP585 estimate of [CO₂] effects, because the estimate does not account for concurrent changes in VPD, drought stress, or possible saturation of WUE under elevated [CO₂]. Consequently, our results represent an attempt to estimate the direct first-order net impacts of climate, vegetation greening, and [CO₂] increase on WY (Zhang et al., 2021b). Future research should adopt more comprehensive models that consider soil-vegetation-atmosphere interactions to better differentiate the contributions of each driving factor to WY.

Third, the improved CCW model does not incorporate certain human activities, such as large-scale irrigation, groundwater pumping, and reservoir regulation, which also limits the assumption that WY = P - ET with negligible storage change. This assumption is mainly appropriate at the annual scale and becomes less reliable in regions where long-term groundwater depletion or reservoir operation substantially alters water storage and runoff routing. For instance, irrigation can sustain vegetation greening during dry seasons, potentially amplifying the vegetation–climate feedback on water yield. Incorporating such anthropogenic processes into the CCW framework through coupled irrigation and water management modules would enable more comprehensive attribution analyses in future studies. Our research also excludes water bodies and built-up land. While urbanization can increase flood risks due to the growing proportion of impervious surfaces (Wasko and Sharma, 2017), these land-use changes represent a small fraction of China's land area.