Evapotranspiration (E) data used in this study were obtained from the Global Land Evaporation Amsterdam Model product (GLEAM v4.2a) (Miralles et al., 2025). This product provides not only total evapotranspiration, but also its major components, including transpiration (Et) and bare-soil evaporation (Eb). It has been widely used in studies of the global water cycle. In GLEAM v4.2a, E is estimated based on the Penman-Monteith framework, in which vapor pressure deficit (VPD) is included as an important atmospheric driver. The main VPD dataset in this study was calculated from monthly 2 m air temperature (Ta) and dew point temperature (Td) from the ERA5-Land reanalysis (Copernicus Climate Change Service, 2019). This choice ensures good consistency with the VPD forcing used in GLEAM, because the VPD input in GLEAM is derived from MSWX Ta and Td, and the MSWX dataset is itself a bias-corrected product partly based on ERA5-Land. In addition to Ta and Td, precipitation (Pre) and surface solar radiation (SR) were also taken from ERA5-Land. Root-zone soil moisture (SM_rz) and surface soil moisture (SM_s) were both obtained from the GLEAM product, in order to keep the soil water variables consistent with the evapotranspiration estimates and component partitioning used in this study. To assess the robustness of the main results, we further introduced one independent VPD dataset and one independent evapotranspiration dataset for cross-validation. The independent VPD dataset was derived from monthly near-surface air temperature (TMP) and vapor pressure (VAP) from CRU TS v4.09, and VPD was calculated following the method described by Xu et al. (2024). The independent evapotranspiration dataset was taken from the P-LSHv2 product, which is a multi-decadal global evapotranspiration dataset with explicit soil moisture constraints (Feng et al., 2025). These two independent datasets were used only for robustness evaluation and cross-validation of the main ERA5-Land- and GLEAM-based results, rather than for the primary analysis itself (Fig. S1 in the Supplement).

Land-use data were obtained from the MCD12C1 version 6.1 product on the Google Earth Engine platform, which provides annual land-cover classification at a spatial resolution of 0.05° (Friedl and Sulla-Menashe, 2022). In this study, the original land-cover classes were further grouped into broader categories for analysis. Leaf area index (LAI) data were taken from the GIMMS LAI V1.2 dataset (Cao et al., 2023a), which has good temporal consistency and has been evaluated against field measurements and Landsat-based LAI samples. Elevation data were obtained from the WorldClim elevation dataset and were used as an auxiliary geographic variable in the analysis (Fick and Hijmans, 2017). According to the aridity index (AI), global land areas were divided into four classes: arid, semi-arid, semi-humid, and humid regions (Lin et al., 2025).

From 1981 to 2020, terrestrial VPD increased markedly at the global scale (0.002 kPa yr⁻¹), but with pronounced spatial heterogeneity (Fig. 1a). The strongest VPD increases were mainly observed in southwestern North America, central Africa, parts of Central Asia, and eastern Australia. Broad increases were also found across northern Eurasia and parts of East Asia (Fig. 1a). The global land monthly climatology showed a pronounced seasonal cycle, with mean VPD increasing from 0.65 kPa in January to a maximum of 1.18 kPa in July, followed by a gradual decline to 0.65 kPa in December (Fig. 1b).

By contrast, terrestrial E showed a more spatially heterogeneous trend pattern (Fig. 1c). Positive trends occurred across 68.9 % of the valid land pixels, while 31.1 % showed negative trends. Increases in E were mainly observed in East Asia, northern Europe, and some high-latitude regions, whereas declines were concentrated in Africa, central South America, southwestern North America, and eastern Australia (Fig. 1c). The monthly climatology of E also exhibited a clear seasonal cycle, with the lowest mean value in February (32.4 mm per month) and the highest value in July (54.6 mm per month) (Fig. 1d). Comparison of the spatial trend patterns shows that some regions with pronounced VPD increases coincided with declining E, whereas E generally increased across many high-latitude regions despite relatively weaker VPD increases.

The apparent sensitivity of terrestrial E to VPD showed pronounced but spatially heterogeneous patterns at the global scale (Fig. 2a). Overall, 60.7 % of land areas exhibited positive apparent sensitivity to VPD, with a global mean sensitivity of 293.3±62.3 mm hPa⁻¹ (mean ± spatial standard deviation). Here, dE/dVPD denotes the moving-window-based apparent local sensitivity of E to VPD, rather than a regression-based interannual response. Spatially, negative sensitivity was concentrated mainly in arid and cold regions, whereas positive sensitivity dominated in relatively warm and humid regions (Fig. 2a). The climatic dependence of dE/dVPD is shown in Fig. 2b. Negative or weak sensitivity occurred mainly under dry climate conditions, especially in warm and low-precipitation bins, whereas positive sensitivity became more widespread in wetter bins and in many cool-to-temperate bins. Overall, the apparent VPD–E relationship varied strongly along temperature and precipitation gradients, but the pattern did not indicate a simple response to temperature alone.

Across land-use types, we grouped the 16 original land-cover categories into four major classes consistent with Figs. 3 and S2: vegetation, grassland, cropland, and barren land. All classes generally showed positive apparent sensitivity to VPD, but with clear differences in magnitude. The vegetation class showed the strongest sensitivity (405.3 mm hPa⁻¹), followed by grassland (342.3 mm hPa⁻¹) and barren land (200.4 mm hPa⁻¹), whereas cropland showed the weakest sensitivity (78.1 mm hPa⁻¹). The latitudinal profiles further show that positive sensitivity was generally greater in vegetation and grassland than in cropland and barren land. Taken together, these results indicate that higher local VPD tended to be associated with higher E in densely vegetated regions under the moving-window estimate, while this apparent relationship was weaker in sparsely vegetated or managed cropland areas.

VPD and total E both showed clear interannual variability during 1981–2020 (Fig. 4a, b). At the global scale, their annual mean series showed broadly consistent interannual variations, especially after 2000. Et showed relatively coherent interannual variations, whereas Eb exhibited stronger fluctuations (Fig. 4c, d).

After controlling for SM_rz and Pre, VPD showed mainly positive partial correlations with Et across global land areas (Fig. 5a). Significant positive correlations (p<0.05) were mainly distributed across the Northern Hemisphere, particularly in North America, Europe, and northern Asia. Positive partial correlations were observed in 65.8 % of valid pixels. Positive correlations also occurred in parts of East Asia, South America, and tropical regions, whereas negative correlations were limited and spatially scattered. Overall, after reducing the effects of water supply, VPD remained positively associated with Et across most land regions, especially in northern extratropical areas. After controlling for SM_s and Pre, VPD showed a weaker relationship with Eb, with a more heterogeneous spatial pattern (Fig. 5b). Significant negative correlations (p<0.05) were mainly concentrated in dry and water-limited regions, including Australia, southern Africa, and parts of central Asia. Negative partial correlations were observed in 59.1 % of valid pixels. Negative correlations also appeared in other low-latitude and subtropical regions, while positive correlations occurred mainly in parts of Europe and East Asia with scattered distributions elsewhere.

The nonlinear relationship between VPD and E varied clearly across AI-based aridity zones (Fig. 6). We compared the one-breakpoint piecewise model, the two-breakpoint piecewise model, and the GAM to identify the dominant transition points in the VPD–E relationship. The one-breakpoint model was used for the main threshold estimates because the two-breakpoint model did not show a clear improvement in model fit across the four aridity zones, based on ΔAIC_c values (Figs. S3 and S6). The GAM curves also showed similar nonlinear transition behavior, which supports the stability of the one-breakpoint estimates.

The estimated thresholds showed a clear aridity gradient. The dominant transition point was highest in the arid zone (1.90 kPa), followed by the semi-arid zone (1.46 kPa). The thresholds were much lower in the semi-humid zone (0.49 kPa) and humid zone (0.47 kPa). These results show that the VPD–E relationship reached its main transition at higher VPD in dry regions but at lower VPD in wetter regions. Therefore, the thresholds reported here should be interpreted as dominant transition points in the observed nonlinear relationship, not as fixed ecological limits. The model comparison among the one-breakpoint model, two-breakpoint model, and GAM was used as a robustness check for the threshold estimates.

The aridity-dependent thresholds suggest that the relationship between VPD and E is jointly regulated by atmospheric water demand and land water supply. VPD can enhance E by increasing atmospheric evaporative demand, but this effect depends on whether sufficient water is available for soil evaporation and plant transpiration. When soil water becomes limited, higher VPD may no longer lead to further increases in E. Instead, it may intensify atmospheric drying and reinforce water limitation by accelerating soil moisture depletion and weakening land–atmosphere moisture feedbacks (Liu et al., 2020).

In arid and semi-arid regions, the higher thresholds indicate that the main transition in the VPD–E relationship tends to occur only under stronger atmospheric demand. This pattern is consistent with the fact that E in drylands is often constrained by water supply before atmospheric demand becomes the dominant control. Therefore, as water limitation intensifies, dry regions may show weak or even negative E responses to increasing VPD. Previous studies have also shown that changes in dryland aridity are controlled by multiple interacting factors, including precipitation, soil moisture, vegetation activity, and atmospheric demand, rather than by VPD alone (Lian et al., 2021). In humid and semi-humid regions, the lower thresholds indicate that the VPD–E relationship can shift at relatively lower VPD levels (Li et al., 2023a). These regions usually have greater water availability, allowing E to respond more directly to atmospheric demand. However, after the transition point, this response may weaken because vegetation regulation, stomatal adjustment, and declining near-surface humidity can limit further water loss. Vegetation responses to climate forcing also vary across hydroclimatic conditions, especially in humid, warm, and forested ecosystems (Gao et al., 2024; Zhou et al., 2019). Thus, the thresholds identified in this study should be interpreted as dominant transition points in the observed VPD–E relationship, rather than as fixed ecological boundaries. Their specific values may vary with vegetation type, soil water storage, climatic background, and the data products used.

The aridity-zone-specific SEM results indicate that VPD is statistically associated with variations in terrestrial E through linked atmospheric, soil-water, vegetation, and energy pathways. This relationship does not only reflect higher atmospheric water demand. It also depends on soil water supply, canopy state, and energy conditions. The SEM structure was built from known ecohydrological processes. The estimated paths were generally consistent with this process basis. VPD was negatively related to SM_rz in all aridity zones, with standardized coefficients of -0.25, -0.61, -0.66, and -0.53 in arid, semi-arid, semi-humid, and humid regions, respectively (Fig. 7a–d). SM_rz and LAI were generally positively related to E. These results suggest that soil water supply and canopy condition are important links between atmospheric drying and actual evapotranspiration. The SEM results should still be interpreted as aridity-zone-scale statistical pathways rather than direct causal effects at individual pixels. Because SEM relies on a predefined directional structure, the estimated pathways describe the relative associations among climatic, hydrological, vegetation, and energy factors within the specified framework, but do not explicitly represent reciprocal feedbacks among evapotranspiration, soil moisture, and atmospheric moisture. The restricted and full SEM comparison also supported the retention of the residual VPD-LAI association in the final model (Table S1).

The direct pathway of VPD to E differed clearly among aridity zones. The effect was strongest and negative in arid regions (standardized coefficient = -0.79; Fig. 7a). It was also strongly negative in semi-arid regions (standardized coefficient = -0.55; Fig. 7b). The effect became weak and non-significant in semi-humid regions (standardized coefficient = -0.06; Fig. 7c). The effect was weaker but still negative in humid regions (standardized coefficient = -0.23; Fig. 7d). This pattern suggests that the remaining direct VPD path may represent statistical associations related to atmospheric drying stress and vegetation regulation after T, Pre, SR, SM_rz, and LAI are included in the model. This path should not be read only as the effect of higher evaporative demand (Wang et al., 2025c). In water-limited regions, higher VPD can increase atmospheric water demand and also increase plant and soil water stress. These combined effects may constrain actual E when soil water availability is limited. Previous studies also reported that E responses to VPD depend on soil moisture, plant water regulation, and background dryness (Massmann et al., 2019; Zhang et al., 2023).

The effect decomposition further supports this climate-dependent mechanism (Fig. S6). The total effect of VPD on E was negative in all aridity zones. The negative effect was strongest in arid and semi-arid regions (standardized total effect = -0.96 and -0.69). The effect was much weaker in semi-humid and humid regions (standardized total effect = -0.06 and -0.14). The overall indirect effect of VPD on E also varied among aridity zones. It was negative in arid, semi-arid, and semi-humid regions (standardized indirect effect = -0.17, -0.14, and -0.01), but it became positive in humid regions (standardized indirect effect = 0.08). This positive indirect effect in humid regions likely reflects compensation from LAI-related and energy-related pathways. LAI had positive effects on E in all aridity zones, especially in semi-humid and humid regions (standardized coefficient = 0.50 and 0.38; Fig. 5c, d). However, the positive remaining relation between VPD and LAI in wetter regions should not be read as evidence that higher VPD directly promotes vegetation growth (Yuan et al., 2019). It more likely reflects remaining co-variation among atmospheric dryness, radiation, temperature, and vegetation activity after soil water and climate factors are controlled (Zhang et al., 2021). Overall, VPD regulation of E is climate dependent. Its main mechanism lies in the joint effects of higher atmospheric water demand, limited soil water supply, canopy state, and energy conditions.

The projected increase in VPD is expected to further influence land–atmosphere interactions by modifying terrestrial E, with potential consequences for regional to global water cycling (Yuan et al., 2019). In the AI-based aridity zones, the threshold analysis shows clear differences in how E may respond under future VPD increase. In the arid and semi-arid regions, the one-breakpoint thresholds are high (1.90 and 1.46 kPa, respectively). This interpretation is consistent with the negative dE/dVPD areas shown in Fig. 2a and with the Aridity-zone-specific SEM results, which showed negative statistical associations between VPD and SM_rz across all aridity zones, especially in semi-arid and semi-humid regions. These results suggest that high VPD may be associated with stronger water limitation in dry regions, which may constrain E.

In the humid and semi-humid regions, the thresholds are much lower (0.47 and 0.49 kPa, respectively). This result suggests that ecosystems in wetter regions may reach the response transition under much lower VPD. These regions may therefore experience earlier response transitions as atmospheric dryness continues to increase, even though E can still rise with VPD over part of the observed range (Fig. 2a). This contrast suggests that future VPD increases may lead to different responses among climate zones, with humid and semi-humid regions potentially reaching response transitions earlier, while dry regions may experience stronger E suppression once water limitation dominates. Such differences may influence ecosystem vulnerability and alter water and carbon cycling, especially in forests and water-limited agricultural systems (Gentine et al., 2019; Will et al., 2013).

Collectively, the sustained rise in VPD may modify land–atmosphere interactions and introduce additional uncertainties into global water and carbon cycling (Gentine et al., 2019). Climate models and risk assessments should therefore better represent the threshold differences among arid, semi-arid, semi-humid, and humid regions. These threshold differences also imply that adaptation measures should be region-specific. Semi-humid and humid regions may need earlier warning because of their lower thresholds, while arid regions require stronger water-management measures under chronic water limitation. Targeted water resource management and ecological conservation measures are therefore needed to strengthen ecosystem resilience and reduce the ecological risks associated with rising VPD.

Data uncertainty remains an important limitation in this study. First, our analysis relies on ERA5-Land reanalysis and global E products. Although ERA5-Land provides broadly consistent land-surface variables, it may still contain uncertainties in soil moisture, especially in arid regions. This uncertainty can affect the inferred VPD–SM_rz–E coupling. This issue is relevant to our SEM because VPD was negatively related to SM_rz across all aridity zones. If soil moisture is biased in dry regions, the estimated strength of the SM_rz-mediated pathway may also be affected. In this case, part of the VPD effect on E may be attributed differently between the direct VPD–E path and the indirect pathway through SM_rz. Second, currently available E products still show substantial discrepancies in long-term means, temporal trends, and spatial patterns. Multi-product assessments have shown that annual E estimates can differ by nearly 50 % across products, with even larger disagreement in regional trends. In addition, the partitioning of total E into transpiration, interception evaporation, and soil evaporation remains highly uncertain among products, which may further affect the estimated magnitude and pathways of VPD regulation. Recent evidence also suggests that biases in E trends can propagate into simulations of near-surface humidity decline and atmospheric drying, implying that E uncertainty can influence not only flux estimates themselves but also the interpretation of land–atmosphere feedbacks. Therefore, the quantitative results reported here should be interpreted with caution, as they remain constrained by differences in input data quality, product algorithms, and E component partitioning schemes. Methodological uncertainty also deserves attention. In this study, threshold estimates describe the main transition in the observed VPD–E relationship but should not be treated as exact ecological thresholds. Their values may be affected by model choice and data uncertainty. In addition, the SEM framework represents directional statistical pathways and does not explicitly capture bidirectional land–atmosphere feedbacks among E, soil moisture, and atmospheric moisture. Future work should use more observations and more flexible nonlinear methods to better constrain the threshold behavior of E under rising VPD.