The Evapotranspiration Stress Index (ESI) (Eq. ) is the ratio of ETa to ETp. Although ETp lacks consensus in definitions and estimation methods, the Penman–Monteith (PM) equation is adopted here due to its consistency with the ESI framework. The ERA5-derived ETp product satisfies this criterion, delivering integrated radiation, surface temperature, and wind speed data required by the PM equation. To avoid potential overestimation of ETp, we implement the Yang et al. (2019)-modified PM formula, which incorporates the CO2 water-saving effect and is widely adopted in evapotranspiration studies (Lian et al., 2021; Zhang et al., 2023). CO2 data were obtained from CarbonTracker (CT2022) for the period 2001–2020 at a 3° × 2° spatial resolution. Appendix  details the calculation procedures of ESI. Additionally, we assess the efficacy of three other mainstream evaporation products in representing ESI spatial patterns across China, benchmarking these against observational data from 26 flux stations. 1ESI=ETaETp

In addition to the meteorological and CO2 data utilized in ETp calculations (Eq. ), total precipitation (Prec), atmospheric pressure (P), and soil volumetric water content (svm) across different soil layers from ERA5 are also considered as influencing factors for the ESI. The root zone of vegetation serves as a critical interface for water–plant–atmosphere interactions (Gao et al., 2014, 2024; Wang-Erlandsson et al., 2016). Utilizing root zone depth estimations based on an 80-year drought return period from Stocker et al. (2023), we computed the spatial distribution of root-zone soil moisture (Rsvm) across China through a weighted averaging algorithm, as detailed in Appendix .

Vegetation factors, key drivers of the Evaporative Stress Index (ESI), were incorporated through the Global Land Surface Satellite (GLASS) datasets, including the Leaf Area Index (LAI; version V60) and Fractional Vegetation Cover (FVC; version V40) (Liang et al., 2013, 2014, 2021). These datasets feature an 8 d temporal resolution and 0.05° spatial resolution spanning 2001–2020. Additionally, Moderate Resolution Imaging Spectroradiometer (MODIS) datasets, including the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) (MOD13C1) with a 16 d temporal resolution, were employed. Besides, to mitigate saturation effects and enhance robustness to noise, kernel density estimation was applied to smooth the NDVI data, creating kernel NDVI (kNDVI) (Camps-Valls et al., 2021) (Eq. ). 2kNDVI=tanh⁡(NDVI2)

Grids with NDVI values below 0.1 are omitted to eliminate barren, rocky, sandy or snow areas. The Aridity Index (AI) is calculated as a multiyear average (2001–2020) by the ratio of annual Prec to ETp using the ERA5 dataset. Lower AI values indicate drier regions, classifying the study area into sub-humid (0.5 ≤ AI < 0.65), semi-arid (0.2 ≤ AI < 0.5), and arid (0.05 ≤ AI < 0.2), with regions exceeding an AI of 0.65 being categorized as non-dryland areas (Spinoni et al., 2015). To mitigate the effects of land-use change, the annual China Land Cover Dataset (Yang and Huang, 2021), featuring a 30 m grid spacing for the years 2001–2020, is employed. Only grids with consistent land cover types (cropland, forest, or grassland) throughout the study duration were selected for analysis, as depicted in Fig. .

The datasets and pre-processing procedures are summarized in Table S1 in the Supplement. All variables were interpolated to a spatial resolution of 0.1° to match the ERA5 dataset and were composited over three distinct temporal scales: 8 d, 16 d, and monthly. The selection of datasets was guided by criteria that included high accuracy, fine spatial resolution, and a broad temporal coverage.

Hydrothermal conditions and vegetation variables can each be represented by multiple proxies. For vegetation health, indicators like FVC, LAI, NDVI, and kNDVI are commonly used. In contrast, the energy conditions considered in this study are limited to T and Rn. A disparity in the number of variables may introduce bias, potentially favouring some feature groups over others, which may distort feature importance assessments (Liu et al., 2024). More critically, a high degree of correlation between features – known as multicollinearity – may cause overfitting, where the model becomes excessively tailored to the training data, compromising its ability to generalize.

To identify the principal factors influencing ET stress, we use Explainable machine learning (SHapley Additive exPlanations), specifically XGBoost and Random Forest (RF) regressors. These models provide a robust approach for feature selection, with SHAP values used to evaluate the importance of each driver, where higher SHAP values indicate greater influence (Lundberg and Lee, 2017). Both XGBoost and RF reduce overfitting risks due to their ensemble nature (Breiman, 2001; Chen and Guestrin, 2016). To optimize the regression model, we use grid search techniques and 5-fold cross-validation for parameter refinement. Furthermore, regularization and the “early stopping” mechanism help manage overfitting and control model complexity. Table S2 in the Supplement presents the performance metrics of two models. Despite the above efforts to prevent overfitting, the relatively low validation accuracy still reveals the SHAP model's susceptibility to misallocating feature importance among highly correlated variables. To address this, we generated multiple training subsets through categorical divisions of temporal scales (8 d, 16 d, monthly) and spatial partitions (drought gradient and land cover types: cropland, forest, grassland). This approach enabled us to obtain diverse importance rankings across China's regions and derive statistically robust importance hierarchies through distribution analysis (Figs.  and S3 in the Supplement).

Moreover, we employ partial dependence plots (PDPs) to elucidate the effects of key drivers on the ESI, which show how the model's predictions change in response to variations in individual drivers, while keeping other variables constant (Friedman, 2001; Štrumbelj and Kononenko, 2014). To enhance the robustness, we performed 30 iterations with varying random seeds and averaged the results.

Persistent common trends and analogous seasonal patterns can result in overfitting, thereby concealing the actual causal relationships (Li et al., 2022a) (Fig. a). In our SHAP model, we input deseasonalized and detrended anomalies of each variable. Seasonality is mitigated by subtracting the climatological mean, and long-term trends are eliminated through the application of a locally weighted smoothing filter (LOWESS) with a span of 0.3 (Cleveland, 1979).

The sensitivity of evaporative stress to key drivers (identified in Sect. ) proxies the ESI's response to driver variations. Given self-amplify mechanisms, the previous state of evaporative stress is modelled by a lag-5 autocorrelation. The mathematical expression for sensitivity is as follows: 3δESIt=θESI1δESIt-1∼θESI5δESIt-5+θHydrδHydrt+θEnerδEnert+θClimδClimt+θVegeδVeget+ε, where δHydrt, δEnert, δClimt, and δVeget are anomalies in hydrological, energy, climatic, and vegetation conditions, respectively, with subscripts indicating that the anomalies are calculated from the current time (t). The corresponding θ represents the sensitivity of evaporative stress investigated in this study, and ε is the error term. θESI1δESIt-1∼θESI5δESIt-5 denote the stress states from the previous 1 to 5 stages. Considering the lagging effect, we constructed a memory dynamic linear model based on Bayesian forward filtering (MDLM) to analyse the time-varying sensitivity of ESI. This model draws upon the foundational work of Liu et al. (2019) and Zhang et al. (2021b), which primarily examines the vegetation's response to antecedent growth states and prevailing climatic conditions.

Dynamic linear models serve as a sophisticated statistical methodology for modelling time-series signals by incorporating various contributing factors (Prado et al., 2021). For each pixel, we developed a MLDM that integrates long-term evaporation stress trends, seasonal cycles, the effects of internal memory from the previous five stages, and external drivers. DLM decomposes the ESI series into three components: local mean and trend, seasonal pattern, and regression part. The MDLM, utilizing Bayesian forward filtering, captures the dynamic relationship between ESI's regression components and driving factors at each time step, providing an “instantaneous” response. Additionally, this approach allows sensitivity analysis in both stationary and non-stationary scenarios (Liu et al., 2019; Zhang et al., 2021b), with model precision increasing as the sample size grows. Therefore, we calibrated MLDM using an 8 d resolution, the finest temporal scale available, with data allocation comprising 80 % for training and 20 % for validation. Further details are available in Text S1 in the Supplement.

We also employed multiple linear regression (MLR) with a 5-year moving window to calculate a rough sensitivity sequence for comparison. To remove seasonality and isolate the relationship between ESI anomalies and drivers, we subtracted the long-term average over an 8 d interval and applied the LOWESS method to eliminate local trends. Using Eq. (), we established a MLR model and analysed the regression coefficients (sensitivity coefficients) for each window. It is crucial to acknowledge that the sensitivity derived from the MLR model represents an average response across a designated window (Fig. c), providing only a rough sequence of sensitivity to confirm the overarching trend throughout the study period.

The regression coefficients obtained by the two methods are partial regression coefficients, reflecting the magnitude and direction of the relationship between the variable and the ESI while holding other independent variables constant to remove the obfuscating effects of collinearity (Toyoda, 2024). The precise elucidation of real-world phenomena's intricacies necessitates incorporating a broad spectrum of representative drivers. To prevent overfitting due to multicollinearity, we evaluated the variance inflation factor (VIF) for the independent variables before modelling (Belsley, 1991) (Table S4 in Supplement).

Besides, we employed Sen's slope estimator (Sen, 1968) to quantify the temporal trend and the Mann–Kendall (MK) test to ascertain the statistical significance (Kendall, 1949; Mann, 1945). The slope > 0 signifies a positive trend, and a slope < 0 denotes a negative trend over time.

This study comprehensively investigates the dynamic trajectory of evaporative stress by comparing the ESI calculated using the conventional PM formula from 1950 to 2020 with the ESI derived from the PM formula incorporating CO2 water-saving effects from 2001 to 2020, enabling a detailed examination of its variation characteristics. Subsequent analyses are conducted using the latter due to limitations in the influencing factors' data series. The SHAP model is applied at 8 d, 16 d, and monthly scales to obtain temporally robust importance rankings. Furthermore, to calculate sensitivity, the MDLM operates at an 8 d temporal resolution, as a larger sample size enhances parameter accuracy. The temporal evolution patterns of sensitivity coefficients are explored at an interannual scale.

Mainland China demonstrated a pronounced decline in the ESI from 1950 to 2020 (total change of 4.74 %, decreasing by 0.54 %d-1, p < 0.001). While the CO2 water-saving effect partially offset the decline, the downward trend remains largely parallel to the long-term series, indicating a continued exacerbation of evaporative stress (-0.70 % per decade, p = 0.31). Annual tracking over 70 years enhances statistical significance due to the expanded sample size (Fig. a). Collectively, the ESI for each month is on a downward trend, with stability in the early growing season (0.58 ± 0.08). Commencing in April, the onset of the growing season prompts a marked increase in ESI and culminates in a peak between 0.9 and 1.0 by August (Fig. b).

Spatially, over one-quarter (28.01 %) of the regions nationwide have registered a significant reduction in ESI (p < 0.05), concentrating in the North China Plain and the Northeast, where the rate surpassing 2.40 % per decade. Concurrently, stress in the Northwest and the middle to lower Yangtze River basin is easing (Fig. c). ESI declines more sharply in dryland (total change of 7.63 %, decreasing by 0.91 % per decade, p < 0.001) than in non-dryland lands (total change of 2.92 %, decreasing by 0.32 % per decade, p < 0.01) (Figs. d and S2 in the Supplement). This decline is observed universally across various land-use categories, with cropland seeing the greatest reduction and the highest percentage of areas (-1.14 % per decade, p < 0.001; 7.59 × 10⁵ km2, 50.24 % of the total cropland area) (Figs. c, e, S1, and S2 in the Supplement). Similar dynamics are observed in ESI factoring in the water conservation benefits attributed to CO2. Even so, the decline in ESI for drought-prone and agriculture-intensified regions over the past 2 decades has outpaced that of the preceding 70 years (-1.63 vs -0.91 % per decade in dryland, -1.60 vs -1.14 % per decade in cropland), highlighting the increasing water stress in these areas.

Both RF and XGBoost models pinpoint shallow soil moisture (0–7 cm) as the paramount driver, indicating that its scarcity is most likely to trigger evaporative stress, especially in water-limited regions, consistent with Liebig's law of the minimum (Danger et al., 2008; Tang and Riley, 2021). Among vegetation proxies, LAI performs the best, likely attributable to its strong correlation with stomatal conductance (Fig. b). Similar importance ranking distributions were observed using RF regressor (Fig. S3b). Comparatively, the XGBoost regressor exhibited enhanced stability across multiple subsets, with minimal divergence between training and validation metrics (Table S2), prompting its selection for subsequent importance quantification analysis. Cross-validation reveals that, across an 8 d interval, water, energy, climate, and vegetation factors on ESI variability decline sequentially, contributing 42 %, 24 %, 21 %, and 11 % to ESI, respectively (Fig. b). Within each category, the drivers exerting the greatest influence are svm0–7 (0.19), Rn (0.18), VPD (0.14), and LAI (0.03) (Fig. a). Monthly-scale analysis reveals amplified hydrological regulation (water contribution = 0.56), with concurrent suppression of vegetation and energy controls (Figs. b, S4, and S5 in the Supplement), demonstrating scale-dependent hierarchy in ecosystem stress controls.

Following collinear variable filtering, we identified svm0–7, Rn, T, VPD, U, P, CO2, and LAI as critical external drivers of evaporative stress (parameter set R0). Given precipitation's secondary ranking to svm0–7, we established scenario R1 substituting precipitation to examine hydrological driver divergence (Appendix ). SHAP model calibration across temporal scales revealed ESI response patterns for both variable sets. Reduced multicollinearity enhanced model stability (Table S3 in the Supplement). ESI generally exhibits a nonlinear relationship with several critical factors (Fig. c–j), decreasing with a higher VPD, T, and Rn but increasing with U, LAI and svm0–7. Similar patterns are observed across temporal resolutions, reinforcing the robustness of our findings (Figs. S4 and S5). Moreover, the non-monotonic interactions between P and CO2 with ESI highlight the imperative for further investigation into how these factors exacerbate or mitigate evaporative stress across various numerical ranges. However, except for Rn, increasing temporal scale generally decreases the linearity and monotonicity of PDPs (Fig. c–j), likely due to data averaging or noise distortion obscuring relationships (Figs. , S4, and S5).

To quantify the influence of these variables on the ESI, denoted by sensitivity parameter θ (Table S5 in the Supplement), factorial simulations are conducted using the MDLM and MLR model. The findings demonstrate solid stability and powerful interpretability after successfully passing the multicollinearity test (Table S4).

Soil moisture stands out as the paramount and most sensitive factor, demonstrating a significant upward trend in both models, albeit with quantitative differences (Fig. b). MDLM results indicate θsvm increased by 9.49 % over the past 2 decades (0.03 ± 0.01 yr-1, p < 0.001), peaking at a 13 % rise from the previous decade in 2016. In contrast, the MLR model, with a slightly lower mean θsvm (2.36 vs. 2.98), exhibited a stronger growth trend (0.05 ± 0.01 yr-1, p < 0.001), peaking earlier in 2014. This shift in peak timing is an inherent consequence of the moving average process. Similarly, Tang et al. (2025) reported analogous evolutionary patterns and attributed this phenomenon primarily to climatic drivers. Among other pivotal drivers, θVPD and θLAI exhibit considerable numerical values and trends (Figs.  and S6 in the Supplement). On average, a 1 kPa rise in VPD correlates with an ∼ 16.84 % rise in evaporative stress according to the MDLM, slightly lower than the 19.74 % increase estimated by the MLR. The MLR also predicts a higher annual growth rate (0.37 ± 0.07 vs 0.24 ± 0.05 %yr-1, p < 0.001) (Fig. c). Regarding LAI, the MLR projects a greater decline, with a cumulative reduction of 82.16 % (compared to 45.77 % in the MDLM) and an annual rate of -0.19 ± 0.02 %yr-1 (vs.-0.16 ± 0.03 %yr-1 in the MDLM, p < 0.05) (Fig. d). This suggests that the perceived decline in vegetation greening efficacy for drought relief, as assessed by conventional MLR methods, might be overstated. The consensus result from Parameter Set R1 corroborates the robustness of this observation (Fig. S6d).

Spatially, θsvm is predominantly higher in the southeast and lower in the northwest, while θPrec exhibits a complementary pattern (Figs. a and S10a in the Supplement). This indicates that in humid regions, ESI is primarily affected by soil moisture regulation, whereas in arid regions, short-term precipitation can substantially alleviate evaporative stress. The specific mechanisms are discussed in Sect. . Additionally, θsvm also varies by land-use type, with forest showing the strongest sensitivity and most drastic changes, followed by cropland and grassland (MLDM: 4.27 > 3.60 > 1.17, Fig. a and d). A plausible explanation is that regions endowed with optimal water conditions and robust ecological structure exhibit increased sensitivity due to the magnified impacts resulting from the disequilibrium of hydrology, climate, and vegetation (Forzieri et al., 2022). Notably, sensitive hotspots were distributed near the “Hu Line” and in the lower Yangtze River basin (Fig. a). Given the recent shifts in forest composition (large-scale planting of single tree species) coupled with the increasing extreme climatic events (Ruan et al., 2022; Yin et al., 2022), these areas warrant special attention. Furthermore, sizeable areas (MDLM: 4.00 × 10⁶ km2, approximately 48.3 % of China mainland) experienced a rise in θsvm (Sen's slope > 0, Fig. a), with a similar pattern observed in MLR (Fig. S9a in the Supplement). However, MLR underestimates θsvm in nearly all vegetated areas (∼ 93.6 %), while markedly overestimating its increasing trend in ∼ 58.5 % of these areas, particularly in the south (Figs. d and S9a).

Similarly, MLR shows a comparable spatial pattern of underestimation in values and overestimation in trends when estimating θVPD (Figs. b, e and S9b). It exhibits notable regional disparities, with dryland encountering a 1.64-fold greater increase in ET stress per unit of VPD compared to wet regions (MDLM: -23.02 % vs -14.07 %). θVPD is decreasing in arid regions (-0.89 % per decade, p < 0.01), in contrast to the increasing trend in humid regions (3.36 % per decade, p < 0.001); therefore the discrepancy between dryland and non-dryland is anticipated to widen (Fig. b and e). Notably, MLR indicates an even greater trend divergence between dryland and non-dryland regions (-2.41 % vs. 6.58 % per decade, p < 0.001, Fig. e). Interestingly, the response of evaporative stress to LAI (θLAI) exhibits in both positive and negative ways. Approximately 29.6 % (2.81 × 10⁶ km2), counterintuitively, experienced an increase in LAI coupled with intensified ET stress (θLAI < 0), mainly in the Loess Plateau, North China Plain, and lower Yangtze River region (Fig. c). In these areas, rapid greening and intensive agricultural expansion are linked to heightened water deficits (Feng et al., 2016; Lu et al., 2018; Wang et al., 2024). In the lower Yangtze River region, increased LAI is associated with reduced evaporation (Zhang et al., 2021c), possibly due to shading effects that lower surface radiation, soil temperature, and soil evaporation (Forzieri et al., 2020). Regions where increased LAI alleviated stress are predominantly near the “Hu Line” (θLAI > 0, 4.01 × 106 km2, ∼ 42.27 %), where sensitivity trends showing a significant decline (Sen's slope < 0, Fig. c). This region is a focal point for China's afforestation efforts (Zhan et al., 2023), implying that the benefits of sustained greening in reducing ET stress might be lessening. Additionally, artificial agronomic regulation renders minimal and stable θLAI in cropland (MLDM: θLAI = 1.11 %, Sen's slope = -0.48 % per decade, Fig. c and f). Figure  further illustrates that the values and trends of θsvm, θVPD, and θLAI, as estimated by MLDM and MLR, display statistically significant differences (p < 0.001) across most climate zones and vegetation types, underscoring substantial methodological discrepancies. Within the same land-use categories, the MLR model consistently demonstrates more dispersed sensitivity trend distributions and higher intra-group variability (Fig. ).

Given the prominent greening tendency in China, we further explored the connection between greening and the sensitivity of dramatically changing variables (Fig. ). In extremely humid regions (AI > 1.5), faster greening trends are associated with lower θsvm values (Fig. a). Conversely, in less humid regions (AI < 1.5), particularly in semi-arid and semi-humid zones (0.2 < AI < 0.65), both θsvm and its trend increase alongside the gradient of greening trends, suggesting higher sensitivity and greater acceleration (Fig. a and b). Specifically, when 0.2 < AI < 0.5, θsvm rises progressively from 1.30 to 2.20, while the associated trend simultaneously escalates from ∼ 0 to 0.2.

Unlike θsvm, θVPD and θLAI (and their trends) do not exhibit a co-varying gradient pattern along the greening trends (Fig. d–i). Remarkably, drier regions with slower pace of greening display greater absolute θVPD values (|θVPD|, Fig. d), suggesting that heightened VPD amplifies ET stress per unit increase. However, in rapid greening areas, |θVPD| tends to be diminished (Fig. d), likely because denser vegetation enhances transpiration, increasing atmospheric moisture and partially offsetting moisture deficit stress (Fig. d and f). These findings imply that vegetation greening may, to some extent, improve the ecosystem's resilience to atmospheric drought. Concerning θLAI, a faster LAI increase corresponds to a steeper decline in θLAI, reducing the efficacy of LAI in mitigating ET pressure (Fig. h and i). Consequently, greening-induced water stress mitigation may be less pronounced than anticipated.

Previous research has largely focused on the long-term average response of individual variables in the hydrological–climatic–vegetation nexus. Our study employs the ESI as a metric to assess the dynamic equilibrium of the system and introduce the MDLM to capture its time-varying response (Fig. ). The MDLM, characterized by its mechanism-driven architecture and operational simplicity, can inherently detrend and deseasonalize data, eliminating the need for anomaly pre-processing. It shows a 44.26 % improvement in simulation accuracy over the moving-window MLR (R2 increasing from 0.61 to 0.88, Figs. a, and S6–S8), consistent with high fitting accuracies (> 0.90) reported in other studies (Zhang et al., 2021b). This improvement is primarily attributed to Bayesian forward filtering embedded in the MDLM for dynamic parameter estimation and its capability to approximate nonlinear dependencies as quasi-linear relationships at discrete temporal nodes (Fig. ). Consequently, it provides time-specific analyses, unlike the window-averaged sensitivity produced by MLR (Liu et al., 2019; Zhang et al., 2022). Pixel-scale model comparisons indicate that the MDLM outperforms the MLR across the entire study area (Figs. S7 and S8 in the Supplement), demonstrating its high flexibility and broad applicability to diverse climatic and terrain conditions. More importantly, comparing the θ sequences for key variables (svm, VPD, and LAI) reveals that the MLR significantly overestimates their temporal trends (Figs. –, S6, S9, and S10). These MLR processes – truncating sequence, smoothing, and simplifying nonlinear dynamics – not only lead to signal loss but also cause substantial differences in θ evolution trajectories (Figs. g, h and S6g). These findings highlight the need to re-evaluate the suitability of traditional moving average methods for determining inter-variable coupling strength or temporal sensitivity trends.

Our findings indicate that despite the water-saving benefits of elevated atmospheric CO2 levels, various regions in China have distinctly experienced increased evaporative stress (Figs. a, b and S1). This phenomenon is in concordance with several recent detrimental environmental shifts, including land surface desiccation, rising evaporation demand, and groundwater depletion (Jasechko et al., 2024; Qing et al., 2023; Yuan et al., 2019). The pronounced increase in ET stress in dryland and cropland underscores the vulnerability of these ecosystems (Figs. c–e, S1, and S2a, c), primarily attributed to the considerable desiccation of soil moisture (Fig. S11a in the Supplement). Soil moisture, as the exclusive direct water supply source for terrestrial evaporation processes – whether from land surface or plant transpiration – exerts a paramount role (Smith and Boers, 2023; Zhao et al., 2023) (Figs. – and S3–S6). Over the past 20 years, increased soil water sensitivity and variability has led to greater ET stress intensification (Figs. b, 7a, and S11a). Sporadic studies on time-varying sensitivity also reveal that water supply increasingly impacts ecosystems (Hu et al., 2023; Li et al., 2022a; Zeng et al., 2022; Zhang et al., 2022). Our simulations, using precipitation as a water proxy for moisture, also corroborated these findings (Figs. S6b and S10). Hence, addressing the threats posed by soil moisture deficits in arid cropland is crucial for ensuring water and food security.

Intriguingly, our research identifies a complementary spatial pattern between θsvm (southern high–northern low) and θPrec across China (Figs. a and S10a). In northern arid/semi-arid regions with low initial soil moisture, precipitation acts as an “instantaneous input” that rapidly replenishes soil moisture and substantially alleviates evaporative stress, exhibiting higher sensitivity. Conversely, in southern humid regions where soils maintain near saturation, surpassing the pre-existing hydro-equilibrium threshold triggers disproportionately severe ecosystem perturbations. Consequently, ESI in these regions demonstrates heightened dependence and sensitivity to soil moisture dynamics. Analogously, despite a relatively moderate increase in forest ET stress (Figs. c–e and S2), soil moisture sensitivity emerges both highly pronounced and rapidly enhanced recently (Fig. a). In regions with favourable hydrological conditions and stable ecological structures, the collapse of the hydrological–climatic–vegetation system can precipitate profoundly adverse effects (Forzieri et al., 2022). Furthermore, in the face of increasingly intense and severe extreme climatic events, complex ecosystems require closer monitoring to formulate proactive strategies.

Additionally, the pace of greening may influence ecosystems' ability to withstand different types of water stress. On the one hand, in semi-arid and semi-humid regions, the sensitivity of ESI to soil moisture and its variability are proportional to the greening trend (Fig. a–c). Therefore, once a soil drought occurs, the adverse impact on ET stress increases, suggesting that overly rapid greening can diminish an ecosystem's resilience to soil drought. Specifically, when afforestation surpasses the carrying capacity dictated by regional hydroclimatic conditions, a slight soil deficit can trigger the self-amplify cycle and result in vegetation degradation and mortality, known as overshoot drought (Zhang et al., 2021b). From a long-term lens, reliance on irrigation during the early stages of artificial vegetation establishment, aimed to ensure plant survival, can hamper the formation of deep root systems (Moreno-Mateos et al., 2020; Xiao et al., 2024b). This practice weakens the plant–groundwater linkage in drylands, subsequently diminishing their resilience (Wang et al., 2023). On the other hand, VPD sensitivity and the greening trend exhibit an inverse gradient (Fig. d), particularly in arid lands with high θVPD values Fig. b), indicating that greening in drylands has improved the ability to cope with atmospheric drought. Two mechanisms may explain this phenomenon:

In certain arid zones, deep-rooted vegetation accesses deep soil moisture during droughts, releasing it into the atmosphere via transpiration, which increases humidity but may also worsen soil aridity (Liang et al., 2024; Sun et al., 2021).

Concerning LAI, a faster greening trend correlates with a diminishing contribution of LAI growth in alleviating water stress (Figs. d, h, and S13h in the Supplement). In certain regions, enhanced LAI may exacerbate water stress (Fig. c). Gleason et al. (2017) found that high vegetation density may exacerbate inter-species competition for water, leading to increased water stress for individual plants or specific communities, providing a cogent explanation for our findings. Collectively, although greening has traditionally been considered a positive environmental adjustment, we advocate for a more critical and dialectical understanding. In adherence to sustainable development imperatives, policymakers and decision-makers are tasked with the precise identification of regions amenable to greening initiatives and the establishment of quantifiable benchmarks that will safeguard ecological equilibrium and foster enduring sustainability.

Although this study provides valuable insights into the interplay between regional greening and complex ecological couplings in China, it remains subject to certain limitations in methodology and data. First, the observational sample size from flux stations is insufficient, as it has been employed solely for the screening of reanalysis products rather than dynamic factorial simulations. Second, deviations in Parameter Set R0 and R1 indicate that other local factors such as soil texture, vegetation acclimation, plant demographic rates, and vegetation–groundwater dependency may exert considerable influence (Abel et al., 2024; Fu et al., 2024b; Patel et al., 2021). More in-depth studies based on long-term ecophysiological observations and purpose-built field experiments are needed to further unravel the complex mechanisms of VPD, soil moisture, and LAI in land–atmosphere interactions.

Finally, what are the key implications given the glimpse of intensifying evapotranspiration stress coinciding with its increasing sensitivity to soil water deficits in a greening China? Soil water variability would lead to greater changes in the ecosystems undergoing rapid greening. That is, stronger drought effects can be expected when soil water is anomalously low. Concurrently, the increase in LAI may not alleviate evaporative stress as much as expected. Combined with the increasing frequency and severity of extreme climate events, a “greening but drying” trend may thus be more prevalent in the future, potentially raising the risk of ecosystem imbalance.