Introduction

Forests absorb roughly 22% of anthropogenic carbon emissions each year, acting as a crucial buffer against climate change1,2. However, the future stability of this buffer may be compromised due to climate change3. The intrinsic water use efficiency (W), which is the ratio of carbon uptake to stomatal conductance for water4,5, is a key physiological metric of trees that underlies the coupling of the carbon and water cycles within forests6,7. Changes in forest tree W can affect terrestrial carbon uptake8 and water use7 and, therefore, have significant implications for forests’ capacity to mitigate climate change.

Observations and models suggest that tree W can be influenced by both biotic (e.g., plant functional types, tree height, and forest ages)7,9,10,11 and abiotic (e.g., atmospheric CO2)7,12,13 factors. For example, tree W is observed to increase with rising atmospheric CO24,7,12, a well-recognized consequence of CO2 fertilization14. Changes in aridity can also impact tree W, with drier conditions leading to enhanced W 15,16,17, a resilience mechanism under water stress. However, it remains uncertain whether these enhancements in W can be sustained over the long term, as recent studies suggest a decrease in the sensitivity of W to rising atmospheric CO2 due to nutrient limitation on carbon uptake18. Few studies have evaluated the impact of shifting water constraints on the variation in W over time. The sensitivity of W to interannual variation in aridity (SW) is critical in determining the drivers of interannual variation in W, especially in the context of climate change15. An increase in SW can lead to a more rapid enhancement in W under the same drying rate, indicating higher drought resilience of trees. Thus, improved spatiotemporal information on SW can provide a mechanistic understanding of the future coupling of carbon and water cycles within forests in response to climate change.

Here, we analyze the interannual relationship between W and Standardized Precipitation Evapotranspiration Index (SPEI) during 1951–2010, leveraging a dataset of isotopic abundance in tree rings and multiple climate datasets. Table 1 provides a detailed explanation of key variables used in the study. The isotopic abundance data are collected from the literature and have been used to study W variation over large regions12,18. We calculate SW as the linear slope of annual W against SPEI at each individual tree using simple linear regression (see “Methods”). As W generally increases under drier conditions (i.e., more negative SPEI), we expect negative slopes for most of the tree sites. Thus, a more negative SW suggests increasing constraint of aridity on W. We then examine the temporal change in SW and evaluate factors controlling its temporal variation over the study period. Specifically, we aim to address two questions: (1) How is the aridity constraint on tree W (i.e., SW) changed between 1951 and 2010? And (2) what factors explain the interannual variation in SW during this period?

Table 1 Key variables used in the manuscript, their corresponding symbols, and units used in the study
Full size table

Results and discussion

Increasing constraint of aridity on tree intrinsic water use efficiency

The detrended annual mean W across all sites (see “Methods”) was negatively correlated with detrended mean SPEI (opposite values presented in Fig. 1a) during 1951–2010 (Fig. 1a). This suggests that a drier environment leads to a higher W, consistent with previous studies13,15,17,19. However, the strength of this association, i.e., the absolute value of the correlation coefficient (R), was not consistent over time (Fig. 1a). In fact, the negative R between W and SPEI over 22-year sliding windows showed a decreasing trend of −0.015 yr-1 (p < 0.001 with a t-test; Fig. 1b) from 1951–1972 (R = −0.40, p > 0.05 with a t-test) to 1989–2010 (R = −0.66, p < 0.001 with a t-test). This decline in R (i.e., more negative R) suggests an enhanced correlation between W and SPEI. The enhanced correlation between W and SPEI is further supported by the similar trend in SW, whose absolute value increased by 112% from 1951–1972 (mean SW = −0.866 μmol mol−1 per unit SPEI, n = 151 sites, p > 0.05 with a one-sample t-test in comparison to 0) to 1989–2010 (mean SW = −1.837 μmol mol-1 per unit SPEI, n = 151 sites, p < 0.001 with a one-sample t-test in comparison to 0) (Fig. 1c). Thus, the strength of the coupling between W and SPEI increased during 1951–2010.

Fig. 1: Change in anomalies of intrinsic water use efficiency (W) and dryness (i.e., standardized precipitation-evapotranspiration index, SPEI), and their correlations during the period from 1950 to 2010.
figure 1

a Interannual variation of W (black) and SPEI (red) anomalies during the period 1950–2010. Reversed values of SPEI anomaly are shown in the figure. b Change in correlation coefficients (R) between W and SPEI anomalies in each 22-year sliding window. c Change in sensitivity of W to SPEI (SW) in each 22-year sliding window. Shaded areas in (b) and (c) indicate 95% confidence intervals in t-tests for each sliding window. Source data are provided as a Source Data file.

Full size image

To further examine the variation in SW across various tree sites, we focused on the change in SW between the earlier period (1951–1972) and the most recent period (1989–2010) at each site (Fig. 2a). This difference estimation can reduce the impact of temporal autocorrelation on SW trend estimation. The majority of sites (91 out of 151) showed a decrease in SW (i.e., an increase in its absolute value) comparing the periods 1989–2010 and 1951-1972 (Fig. 2b), with a mean change of −0.379 μmol mol−1 per unit SPEI (n = 151 sites, p < 0.05 with a paired t-test). Meanwhile, cross-site variations in the temporal change of SW did not show a dependence on site-specific climatic conditions (Fig. 2c) such as mean annual temperature (MAT: −13.8–27.1 °C, R = 0.00, p > 0.1 with a t-test) and mean annual precipitation (MAP: 101–2766 mm, R = 0.16, p > 0.05 with a t-test). We also found similar relationships when using growing season temperature (MATGS) and precipitation (MAPGS) instead (Supplementary Figs. 1 and 2).

Fig. 2: Change in sensitivity (SW) of intrinsic water use efficiency (W) to standardized precipitation-evapotranspiration index (SPEI) between the two periods 1951–1972 and 1989–2010.
figure 2

a Spatial distribution of the SW change from individual tree sites. The background world map is obtained from the publicly available “maps” package in R. b Probability density of the SW change from individual sites. c, Distribution of the SW change under pairs of mean annual temperature (MAT) and mean annual precipitation (MAP) at sites. In (a, c), the triangle and circle points indicate angiosperms (n = 42 sites) and gymnosperms (n = 109 sites), respectively. Darker colors indicate larger changes in SW. Source data are provided as a Source Data file.

Full size image

Empirical explanation for the increased sensitivity of intrinsic water use efficiency to aridity variation

To account for divergent water use strategies across different clades9,20, we compared the change in SW between angiosperms and gymnosperms. Gymnosperms exhibited a decrease in SW (i.e., an increase in its absolute value) at over half of the sites analyzed (62%; 68 out of 109 sites) from the period 1951–1972 to 1989–2010. On average, SW of gymnosperms declined by 0.524 μmol mol−1 per unit SPEI across all the 109 sites (p < 0.01 with a paired t-test; Fig. 3a). By contrast, angiosperms did not show statistically significant changes in SW (Fig. 3a). These differences may arise due to contrasting plant functional traits (e.g., water potentials at which 50% of hydraulic conductivity is lost, hydraulic safety margin, leaf type, and specific leaf area) found among angiosperms and gymnosperms, which determine their diverse responses to environmental changes, including water availability20,21,22,23.

Fig. 3: Comparing changes in sensitivity (SW) of intrinsic water use efficiency (W) to standardized precipitation-evapotranspiration index (SPEI) between 1951–1972 and 1989–2010 across clades and aridity trends.
figure 3

a Variation in SW changes between angiosperms and gymnosperms. The triangle and horizontal bar in the box indicate the mean and median of the change in SW, respectively. The sample size is 42 and 109 for angiosperms and gymnosperms, respectively. In each boxplot, the upper and lower boundaries of the box represent 75th and 25th percentiles, respectively; the upper and lower whiskers represent a distance of 1.5 times the interquartile range (IQR) above the 75th percentile and below the 25th percentile, respectively. The IQR indicates the distance between the 75th and 25th percentiles. The points beyond the whisker boundaries are potential outliers. The asterisk mark above the box indicates the mean value lower than 0 with a significance level of 0.05 based on a one-sample t-test. b The change in SW against that in AI for angiosperms. Each triangle indicates one angiosperm site. c The change in SW against that in AI for gymnosperms. Each circle indicates one gymnosperm site. In (b and c), the dark (light) colors indicate negative (positive) changes in AI, and the brown (green) colors indicate negative (positive) changes in SW. Source data are provided as a Source Data file.

Full size image

We suppose that climate change may also have contributed to the more negative SW, especially in gymnosperms, during 1989–2010 compared to 1951–1972. We noticed an overall decrease in mean SPEI (p < 0.001 with a paired t-test) across sites (n = 151), indicating a drying trend due to climate warming from 1951–2010 (Supplementary Fig. 3). This trend was further confirmed by the decline in the Aridity Index (AI), which is the ratio of precipitation to potential evapotranspiration (see “Methods”), between the two periods (Fig. 3b, c). The decline in AI between the periods was −0.026 (p < 0.001 with a paired t-test; negative values in 68 out of 109 sites) across the gymnosperm sites. With climate change intensifying drought in many regions since 1950s24,25, trees have become more vulnerable to drought-induced damage26. In response, trees tend to rapidly close stomata to prevent further damage to their hydraulic system27,28, which may contribute to their increased sensitivity to subsequent droughts, especially for gymnosperm species29. Under drier conditions, faster stomatal closure in response to increased aridity can make trees more conservative in their water use, leading to a larger change in W for a given level of dryness intensity (i.e., higher SW), all else being equal30. Note that gymnosperm sites with both negative changes in SW and AI represent the largest proportion (~35%).

However, it is important to exercise caution when interpreting changes in SW at individual sites across heterogeneous environments (Fig. 3). For example, about 30 out of 68 gymnosperm sites showed a negative change in SW but no decreases in AI (Fig. 3c). This suggests that, at some sites, factors other than increasing dryness could have played a more dominant role in influencing SW variation over time. One potential factor is the long-term rise in atmospheric CO2 concentrations, which may partially explain the observed negative trends in SW. Indeed, rising CO2 can drive a suite of physiological adjustments in plants, including alterations in hydraulic (e.g., hydraulic conductivity) and leaf structural (e.g., leaf mass per area) traits31,32, maximum photosynthetic capacity (e.g., photosynthetic carboxylation capacity)33, and stomatal gas exchange behaviors34,35. These changes can, in turn, influence the response of W to changes in dryness (i.e., SW). However, results from lab studies and Free-Air CO2 Enrichment (FACE) experiments are sometimes inconsistent regarding the effects of elevated CO2 on stomatal conductance5,35. This discrepancy may arise from differences in experimental design, species studied, and environmental controls, highlighting the challenge of extrapolating short-term experimental findings to long-term field responses14.

Although diverse stomatal conductance responses to elevated CO2 have been observed35, trees consistently exhibit photosynthetic stimulation and acclimation14. Meanwhile, leaf hydraulic conductivity has been found to decline under elevated CO2, which is correlated with an increase in leaf sensitivity to cavitation under drought31. However, it is challenging to directly quantify the effect of CO2 on temporal variation in SW through empirical observations alone. One approach to distinguish the contribution of increased atmospheric CO2 is by conducting controlled field experiments (e.g., FACE experiments). Another approach, as we have adopted, is through modeling.

Theoretical mechanisms for the increased sensitivity of intrinsic water use efficiency to aridity variation

To better understand how dryness and atmospheric CO2 (Ca) influence SW, we used a theoretical model to simulate SW variation under different conditions36,37 (see “Methods”). Since the model does not explicitly include SPEI as an input variable, we used soil moisture (SM) in the model as a proxy for SPEI. While this substitution may bias quantitative estimations of theoretical SW changes, the positive linear correlation between annual SM and SPEI (median and mean correlation coefficient: 0.75 and 0.72, respectively; Supplementary Fig. 4) supports using SM to qualitatively evaluate SW variation under various factors (see “Methods” for details). The simulations revealed that W generally becomes more sensitive to dryness variation (i.e., more negative SW) as Ca increases and SM decreases (Fig. 4a). Furthermore, decreased SM and increased Ca amplify each other’s effect on SW changes (Supplementary Fig. 5). When considering stomatal slope parameter (g1) acclimation to increased Ca38 and climate change39,40 in simulations (see “Methods”), results are generally consistent (Supplementary Figs. 6 and 7). However, under expected g1 acclimation rate to global environmental change, decreased SM and increased Ca diminish each other’s effect on SW changes (Supplementary Fig. 7). This discrepancy does not negate the overall impact of SM and Ca on SW changes but suggests that g1 acclimation may counteract these environmental effects under extreme dry conditions (Supplementary Fig. 7b). Meanwhile, we advise caution in interpreting results involving g1 acclimation, as its dynamics remain poorly constrained39,41. Further research is needed to clarify how g1 adapts over time to global environmental changes and to incorporate these dynamics into models. In summary, the modeling results for SW under simulated drier conditions with increased Ca qualitatively align with our empirical observations, despite potential biotic acclimation effects.

Fig. 4: Simulation of the sensitivity (SW) of intrinsic water use efficiency to soil moisture (μmol mol−1 kg−1 m2) based on a theoretical model.
figure 4

a Simulated SW under theoretical changes in atmospheric CO2 (Ca) and soil moisture (SM). Ca ranges from 300 to 400 μmol mol−1, and SM ranges from 10 to 50 kg m−2. b Simulated SW under different scenarios derived from Coupled Model Intercomparison Project Phase 6 (CMIP6) models. The “historical” scenario represents simulations for the period 1989–2010. Other scenarios represent simulations for 2029–2050 under different Shared Socioeconomic Pathways (SSPs). Each point represents a simulated SW from one model. The sample size for each scenario (i.e., “historical”, “ssp126”, “ssp245”, “ssp370”, and “ssp585”) is 13, 10, 9, 9, and 10, respectively. In each boxplot, the upper and lower boundaries of the box represent 75th and 25th percentiles, respectively; the upper and lower whiskers represent a distance of 1.5 times the interquartile range (IQR) above the 75th percentile and below the 25th percentile, respectively. The IQR indicates the distance between the 75th and 25th percentiles. The horizontal line in each box indicates the median value, while the dashed line indicates the mean SW of historical simulations. Source data are provided as a Source Data file.

Full size image

However, the model simulation cannot explain the non-significant changes in SW for angiosperms in the study (Fig. 3). This limitation may stem from the model’s simplifications or from the distinct characteristics of angiosperms compared to gymnosperms20. Angiosperms adopt riskier water use strategies to cope with water stress42, leveraging diverse physiological pathways to adjust their responses to lower water availability. These strategies include embolism refilling, re-sprouting from branch nodes, rapid leaf area adjustments, and substantial storage capacities for water and nonstructural carbohydrates20,43. While some of these mechanisms remain debated44, this diverse portfolio of strategies may alleviate aridity constraints on angiosperms’ physiological processes, such as stomatal conductance. Furthermore, increased atmospheric CO2 could also mitigate the adverse effects of reduced water availability. Together, these factors may contribute to the reduced temporal variability of SW in angiosperms.

Considering four commonly used future scenarios from the Shared Socioeconomic Pathways (SSPs)—which represent different trajectories of greenhouse gas emissions and socioeconomic development (Riahi et al. 2017)—namely SSP126, SSP245, SSP370, and SSP58545, the theoretical model using data from multiple Earth system models projects increasingly negative mean SW values (Table S1), declining by 0.43, 0.97, 0.70, and 1.11 μmol mol-1 kg-1 m2, respectively (Fig. 4b). These declines are not linearly related to changes in atmospheric CO2 and its radiative forcing implied by the SSPs. This suggests that future aridity constraints on the intrinsic water use efficiency of trees are likely to intensify in a complex manner, driven by multiple environmental factors. Taken together, these findings shed light on how SW changed during 1951–2010. The more sensitive responses of W to dryness variation, especially for gymnosperms, are associated with the drier climate and rising atmospheric CO2 over time. This implies a resilient response of tree intrinsic water use efficiency to aridity under global environmental changes. Such resilience may mitigate the impacts of drought intensification on tree growth and survival in the context of global warming46,47, though only up to a certain threshold.

Our study has revealed that tree water use efficiency (W) has become increasingly sensitive to dryness variation from 1951 to 2010, resulting in an increasing control of aridity on tree and forest functioning. The adaptive response of W to aridity is tree-clade dependent, statistically significant for gymnosperms but not for angiosperms. This change can be attributed to both a drier environment with climate change and rising atmospheric CO2. The amplified SW signifies stronger water constraints on tree intrinsic water use efficiency over large regions, highlighting the need for a better understanding of tree resilience to drought. In future climate change scenarios, shifts in forest species compositions and community dynamics in response to intensified drought48 may also increase ecosystem-level W and SW, ultimately altering carbon and water exchanges49. Our study highlights the importance of considering plasticity and adaptation of W for better predictions of future carbon and water cycles.

Methods

Tree ring stable carbon isotopes

Tree ring stable carbon isotopes (δ13C) collected in Adams et al.18 were used in the study to estimate tree intrinsic water use efficiency (W). The dataset includes 425 individual trees (147 angiosperms and 278 gymnosperms) with 136 species. Each tree contains original data from literature which are one of data types below: δ13C, capital delta (Δ), W, and intercellular CO2 concentrations (ci). Either of these data is an annual time series between 1850 and 2015. Note the original tree-ring δ13C value has been adjusted to a pre-industrial atmospheric stable carbon isotope ratio standard of −6.4‰18,50. Supporting information for each tree, e.g., species name, latitude, and longitude, is also included in the dataset. Focused on the period 1951–2010, we finally chose 296 sites for the further analysis (see “Empirical analysis”)

Atmospheric CO2 concentration and stable carbon isotopes

Atmospheric CO2 concentration (ca) and its stable carbon isotope ratio (δ13Ca) were used to reconstruct annual W for each tree. We used the atmospheric CO2 concentration and stable carbon isotopes recommended by Belmecheri and Lavergne51. Both the datasets consider the inter-hemispheric gradient of the variables. The ca data were produced by compiling published ice core records and recent measurements of firn air and atmospheric samples. The δ13Ca data were a compilation of ice core and firn δ13C records updated to 2016 CE and δ13Ca measurements from Mauna Loa Observatory in Hawaii and the South Pole. In this study, data from the period 1850–2015 were used.

Climatic data

We used the Climate Research Unit (CRU) TS (v4.04) dataset52 to calculate the climatic conditions of the tree sites. The climatic conditions include mean annual temperature (MAT), mean annual precipitation (MAP), and aridity index (i.e., AI). We selected global gridded monthly temperature, precipitation, and potential evapotranspiration from the CRU TS (v4.04) dataset. The dataset spans from January 1901 to December 2019 with a spatial resolution of 0.5 × 0.5 degree. To estimate the MAT at each site, all the monthly temperature data during the period 1951–2015 were averaged into a mean. To estimate the MAP at each site, the monthly precipitation data at each year were first summed up to an annual value and then averaged over the period 1951–2015. The AI was estimated by a ratio of annual precipitation to potential evapotranspiration. Similar to the estimation for the annual precipitation, the annual potential evapotranspiration was calculated by summing up the monthly data at each year. Then the annual AI data during the defined period (see “Empirical analysis”) were averaged to indicate the dryness condition of the site.

We used the standardized precipitation-evapotranspiration index (SPEI) to estimate the interannual variation in dryness of each site. The SPEI was obtained from the global gridded SPEI dataset (SPEIbase v.2.6), which spans from 1901 to 2015 with a spatial resolution of 0.5 × 0.5 degree53,54. The growing-season mean SPEI was estimated from the monthly data at the 6-month scale from April to October. We then detrended the growing-season mean SPEI (i.e., SPEIIAV) by subtracting the linear trend of the SPEI time series estimated from the simple linear regression.

We used the soil moisture (SM) derived from the Global Land Data Assimilation System (version 2.0; GLDAS-2.0) dataset to indicate that SM can be a proxy for SPEI at individual tree sites (Supplementary Fig. 4). The GLDAS-2.0 products were generated by using advanced land surface model (i.e., Noah) forced by the Princeton global meteorological forcing data55. The dataset spans from January 1948 to December 2014 with a spatial resolution of 0.25 × 0.25 degree at a monthly scale, generally covering the period (1951–2015) our study focused. We extracted the monthly data at each tree site according to the latitude and longitude and averaged the growing-season values (i.e., April–October) in each year to obtain the annual mean.

Water use efficiency (W) estimation

For each tree, the W was estimated based on the stable carbon isotope data from the tree ring dataset. The calculation follows the equation below:

$$W={c}_{a}\left(1-{c}_{i}/{c}_{a}\right){\mbox{/}}1.6$$
(1)

where ca and ci are the ambient atmospheric and intercellular CO2 concentrations (μmol mol-1), respectively. The ratio of ci to ca can be estimated by the method in Keeling, et al.56, which considered the effects of mesophyll conductance and photorespiration:

$${c}_{i}/{c}_{a}=\left(\Delta -a+\left(b-{a}_{m}\right)\left(A/{c}_{a}\right)/{g}_{m}+f\left({\varGamma }^{*}/{c}_{a}\right)\right)/\left(b-a\right)$$
(2)

where Δ is the measurement of the carbon isotope discrimination by the tree; a (4.4‰) and b (30‰) are the discrimination coefficients for stomatal diffusion and carbon fixation by Ribulose-1,5-bisphosphate carboxylase-oxygenase (Rubisco), respectively; am is the fractionation associated with internal CO2 transfer within the mesophyll (1.8‰); A is the photosynthetic rate (μmol m−2 s−1); gm is the mesophyll conductance to CO2 taking the typical value 0.2 (mol m−2 s−1); f is the discrimination coefficient for the photorespiration (12‰); Γ* is the CO2 compensation point in the absence of mitochondrial respiration taking the typical value 43 μmol mol−1)18. It is worth noting two important aspects of the analysis. (1) While different values of the parameter b (e.g., 27‰ or 28‰) can lead to variations in ci/ca (Supplementary Fig. 8), the ci/ca values remain highly correlated throughout the study period when examined at a single site (Supplementary Fig. 9). This strong correlation ensures that the sensitivity estimation is not undermined. (2) We accounted for mesophyll conductance in our estimation because previous studies have demonstrated that including this factor significantly improves the accuracy of W estimation when using carbon stable isotopes57.

We calculated the Δ by correcting the stable carbon isotope in woods for the decreased isotopic composition of atmospheric CO2:

$$\Delta=({\delta }^{13}{{{\rm{C}}}}_{{{\rm{a}}}}-{\delta }^{13}{{{\rm{C}}}}_{{{\rm{w}}}})/(1+{\delta }^{13}{{{\rm{C}}}}_{{{\rm{w}}}}/1000)$$
(3)

where δ13Ca and δ13Cw are the stable carbon isotope ratio of atmospheric CO2 and wood cellulose (‰), respectively.

We calculated the A/ca directly following the method in Adams, et al.18 which used a typical equation describing the variations in A/ca:

$$A/{c}_{a}={\left(\frac{A}{{c}_{a}}\right)}_{280}{\left({c}_{a}/280\right)}^{\beta }$$
(4)

where (A/ca)280 is the pre-industrial value of the ratio at which A and ca equal to 9 (μmol m−2 s−1) and 280 (μmol mol−1), respectively; β represents the equation \({\mathrm{ln}}\left({\mbox{DR}}\right)/{\mathrm{ln}}\left(2\right)-1\), where DR is the doubling ratio indicating the proportion of A with a doubling of ca compared to A at the original ca. Since A increases by 45% when ca doubles from 280 ppm to 560 ppm, DR is set to 1.4518,56.

Sensitivity of water use efficiency (W) to aridity variation

The sensitivity of W to aridity variation (SW) is defined as the change in W per unit change in aridity (i.e., SPEI). A linear regression analysis58 of the W against the SPEI is used to estimate SW. SW is defined as the slope coefficient of the linear regression model. A negative coefficient (i.e., negative SW) is expected for most of the tree sites, since W usually increases when the water condition gets drier (i.e., lower SPEI).

At each site, we first de-trended the W time series by using a “cubic smoothing spline” approach which is provided in the “dplR” package in R language and environment59. The method is supposed to remove non-climatic signals (e.g., increased CO2, Suess effect for δ13Ca, and tree growth effect) with a 50% frequency-response cut-off at 20 years from the time series60,61. Subtracting a low-frequency component from the original W signal leads to the interannual variation of the W 60. The SW was then estimated by using the simple linear regression (i.e., the temporal model):

$${W}_{{\mbox{IAV}}}={a}_{0}+{a}_{1}{{\mbox{SPEI}}}_{{\mbox{IAV}}}+\varepsilon$$
(5)

where the WIAV is the interannual variation of the W; a0 is the interception of the simple linear model (μmol mol-1); a1 is the slope of the simple linear model representing the SW (μmol mol-1 per unit SPEI); SPEIIAV is the interannual variation of the detrended growing-season SPEI; ε is the residuals of the model.

Empirical analysis

To examine the temporal variation in the coupling between W and aridity, we calculated the correlation coefficient (R) between W and SPEI in a 22-year window through 1901–2015. Both W and SPEI were first detrended by subtracting their linear trends. Since W is generally negatively correlated with SPEI, the more negative value of R indicates the higher coupling between the two variables. R is estimated by the “Pearson” method.

To investigate the temporal variation in SW, we estimated the SW in a 22-year sliding window through the period 1901–2015 for the tree ring data at each site. The length of the sliding window (i.e., 22) can deliver a robust estimation for the SW variation. We also calculated the SW by using the window size of 20, which gives us consistent results. In each window, the simple linear regression was used to derive the SW at each site. The SW was estimated only if the number of observations in the window is no less than half of the window size (i.e., > 11 years) so that enough data can be used to obtain a robust estimation. We constructed two intervals: 1951–1972 and 1989–2010, during which fossil fuels became the dominant source of emissions2, and selected 151 trees (42 angiosperms and 109 gymnosperms) all of which contain the estimated SW in both periods. We used the t-test to examine if there is any significant change (p < 0.05) in SW between the two periods. To illustrate the factors associated with the change in SW, we first estimated mean annual AI in the window which corresponds to the two intervals (i.e., 1951–1972 and 1989–2010) at each site. The change in the mean annual AI, which indicates dryness the sites suffered during the period, was calculated as the difference of AI values between the two periods.

Analytical derivation of sensitivity of W to aridity variation

We used a stomatal model derived from optimality principles to simulate W variations as influencing factors (i.e., soil moisture and CO2) change to illustrate impact of environmental changes on the temporal dynamics of the SW. These theoretical simulations can help reveal primary mechanisms behind the non-linear change in the W response to aridity variation. The model, derived from Medlyn, et al.36, assumes no residual stomatal conductance (g0) at the light compensation point, i.e., g0 is set to 036:

$$W=\frac{\sqrt{D}\cdot {c}_{a}}{1.6\left(\sqrt{D}+{g}_{1}\cdot \beta \right)}$$
(6)

where D is vapor pressure deficit (kPa); ca is ambient atmospheric CO2 concentration (μmol mol-1); g1 is a slope parameter (kPa0.5) that varies with species and environmental conditions30; β is a function of soil moisture describing soil water stress. We used the β function as described below which is similar to what Sabot et al. did for their control model37:

$$\beta={\left(\frac{{\mbox{SM}}-{{\mbox{SM}}}_{w}}{{{\mbox{SM}}}_{c}-{{\mbox{SM}}}_{w}}\right)}^{a}$$
(7)

where SM is volumetric soil moisture (kg m-2); SMw and SMc are volumetric soil moistures (kg m-2) at the wilting point and the field capacity, respectively; a is a unitless power factor which is set to 0.5 without losing the generality (Supplementary Fig. 10). The curve of β against SM is consistent with previous results from field measurements62, suggesting that the parameter setting is reasonable.

The model does not directly use SPEI as an input; instead, it incorporates D and SM as dryness variables. We first show that interannual variations in SPEI exhibit a strong, positive linear correlation with those in SM at most sites (mean of the correlation coefficients: 0.72, p < 0.001; Supplementary Fig. 4). This linear correlation suggests that changes in SW can theoretically be represented by those in the sensitivity of W to SM (i.e., SW_SM), despite their different units. Thus, in the theoretical analysis, we used SW_SM to infer the behavior of SW under different levels of ca and dryness. To facilitate the qualitative analysis of potential mechanisms driving changes in SW, we excluded the dD term from the model (see derivations below), as D does not show a consistent variability with SPEI. However, for quantifying the contributions of different factors to SW changes in future research, the dD term must be included. Based on the theoretical model, we derive the partial derivation of W to the variables representing the theoretical sensitivity of:

$${dW}= \left\{\frac{\sqrt{D}}{1.6\left(\sqrt{D}+{g}_{1}\cdot \beta \right)}\right\}d{c}_{a}+\left\{\frac{{c}_{a}\cdot {g}_{1}\cdot \beta }{2\sqrt{D}}\cdot \frac{1}{1.6{\left(\sqrt{D}+{g}_{1}\cdot \beta \right)}^{2}}\right\}{dD} \\ + \left\{-\frac{\sqrt{D}\cdot {c}_{a}}{1.6{\left(\sqrt{D}+{g}_{1}\cdot \beta \right)}^{2}}\cdot a\cdot {g}_{1}\cdot {\left(\frac{{\mbox{SM}}-{{\mbox{SM}}}_{w}}{{{\mbox{SM}}}_{c}-{{\mbox{SM}}}_{w}}\right)}^{a}{\left({\mbox{SM}}-{{\mbox{SM}}}_{w}\right)}^{-1}\right\}{dSM}$$
(8)

where dW, dca, dD, and dSM represent the changes in W, ca, D, and SM, respectively. The SW_SM is the term in the curly brackets before dSM. In the simulations, we set the parameters as shown below: SMw = 5 kg m−2, and SMc = 80 kg m−2. g1 was set 2.35 (kPa−2) which is a typical value for gymnosperms30. The other variables are generally consistent with the growing-season values observed at the tree sites during the period 1951–2010: D was set to a range from 0.1 to 3.5 kPa with a 0.1 interval while SM changes from 10 to 50 kg m−2, inversely corresponding to the D values. ca ranges from 300 to 400 μmol mol−1.

Simulations considering g 1 acclimation

Acclimation of g1 to atmospheric CO2. We simulated SW, accounting for g1 acclimation to atmospheric CO2 (Ca) based on relationships identified by Gardner et al.38. Using their data, we estimated the sensitivity of g1 to Ca as 1/600 kPa0.5 ppm−1. The g1 value at 300 ppm Ca (2.35 kPa0.5) was linearly extrapolated to values within the Ca range of 300 to 400 ppm.

Acclimation of g1 to changes in other environmental factors. We adopted the numerical experiment design from Mastrotheodoros et al.39 to assess the impact of g1 acclimation on changes in SW. Specifically, g1 was assigned a decreasing trend of 1% per year over 60 years (1951–2010), with the initial value set at 2.35 kPa0.5. Since g1 reflects stomatal sensitivity to environmental conditions, its decrease over time aligns with expected plant responses to climate and global change. Specifically, rising atmospheric CO2 and increasing aridity can lead to tighter stomatal regulation, reducing g1 as plants optimize water use and maintain carbon uptake under drier or more CO2-rich conditions36. Atmospheric CO2 concentrations were increased from 300 to 400 μmol mol−1, while soil moisture (SM) ranged from 50 to 10 kg m−2, and vapor pressure deficit (D) ranged from 0.5 to 3.5 kPa0.5, consistent with previous settings.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.