Introduction

Drylands cover ~41% of global terrestrial areas and provide crucial ecosystem services. However, ecosystems in drylands have been altered by noticeable climate change, coinciding with substantial efforts of ecological restoration1. As a nature-based solution for climate change mitigation2, ecological restoration has been conducted across millions of hectares of drylands worldwide over the past few decades3. Changes in the land surface are tightly coupled with the dynamic of climate state through land‒atmosphere coupling4,5,6. For instance, ecological restoration has been shown to provide the benefit of alleviating anthropogenic warming7,8,9.

Intense land‒atmosphere coupling is consistently observed in ecological transitions areas10,11,12. It not only contributes to hotter and drier climate conditions in the context of global warming13,14 but also promotes a tight connection between heatwaves and droughts, causing them to amplify each other rapidly15,16,17,18,19,20,21 and propagate over time and space22. The exacerbated concurrent soil drought and atmospheric aridity have caused an irreversible tree dieback in drylands23,24,25,26,27,28. To restore degraded ecosystems, countries around the world are actively implementing ecological restoration programs, and ecological restoration has influenced local thermal and hydrological conditions through land‒atmosphere coupling29,30,31,32,33,34,35. Furthermore, changes in the land surface following ecological restoration, such as the excessive consumption in evapotranspiration which subsequently reduces soil moisture36,37,38,39,40, can potentially lead to a modified land‒atmosphere coupling29,41,42. There are distinct roles for vegetation and soil in land‒atmosphere coupling. Greening alleviates the increase in temperature and vapor pressure deficit (VPD) through vegetation‒atmosphere coupling34,35, while soil drying aggravates the increase in temperature and VPD through soil moisture‒atmosphere coupling43. It remains unknown whether ecological restoration potentially mitigates atmospheric drought, and what the different roles of changes in vegetation and soil moisture are in regulating atmospheric drought, given that the thermal and hydrological processes of plants and soil moisture are arguably far from independent36.

In this study, we focus on VPD, a key factor of atmospheric drought threatening ecosystem stability in drylands44,45, and quantify the impact of ecological restoration on mitigating VPD through land‒atmosphere coupling. We choose China’s drylands as our study area because of the well-known intensive land‒atmosphere coupling10 and the widely reported extensive ecological restoration of returning farmland to grassland and woodland at the end of 199032. We focus on the period from May to September, characterized by significant moisture and energy exchange between the land surface and the atmosphere in China’s drylands46. We combine the use of 1-month lag correlation and the structural equation model (SEM) to separate the pathways of land‒atmosphere coupling derived from vegetation and soil moisture. The strength of land‒atmosphere coupling is quantitatively assessed by examining the path coefficient between leaf area index (LAI)/soil moisture and 1-month lag VPD (VPD+1), considering that the optimal lag time is 1 month47. Furthermore, we quantify the contribution of two pathways (vegetation‒VPD coupling and soil moisture‒VPD coupling) to changes in VPD within the soil-vegetation-atmosphere continuum, by decomposing the effects of each current-month’s variables (including climate conditions, vegetation, and soil moisture) on VPD+1 in SEM. Finally, moisture and energy exchange between the land surface and the atmosphere is explored to provide the mechanistic explanation of changes in land‒atmosphere coupling using machine learning methods. Our study aims to lead a comprehensive understanding of the influence of ecological restoration on land‒atmosphere coupling and provide insights for guiding ecological restoration in drylands in China and worldwide.

Results

Land‒atmosphere coupling following the aridity gradient

Land surface variations (including vegetation and soil moisture) are closely related to the following climatic variables, which can be explained by land‒atmosphere coupling (Supplementary Note 1). Here we analyze path coefficients between the land surface and the atmosphere in SEM, and find that LAI is negatively related to VPD+1 in areas where AI ≤ 0.65 (Fig. 1b), indicating that VPD is widely mitigated by the increase in vegetation transpiration (Fig. 1a). Notably, strong negative effects occurred in areas where 0.3 < AI ≤ 0.5, the path coefficient between LAI and VPD+1 (LAI-VPD+1) are −0.10 in areas where 0.3 < AI ≤ 0.4 (p < 0.05) and −0.11 in areas where 0.4 < AI ≤ 0.5 (p < 0.05). Soil moisture is also negatively related to VPD+1, indicating that the VPD increase is amplified by the decrease in soil evaporation (Fig. 1a). The intensive negative effects occurring in areas where 0.4 < AI ≤ 0.65, the path coefficient between soil moisture and VPD+1 (SM-VPD+1) are −0.39 in areas where 0.4 < AI ≤ 0.5 (p < 0.05) and −0.31 in areas where 0.5 < AI ≤ 0.65 (p < 0.05; Fig. 1b).

Fig. 1: Spatial pattern of the strength of land‒atmosphere coupling.
figure 1

a Changes in LAI (green circle) and soil moisture (brown rectangle) during 1982–2018 across the aridity gradient. b Coefficient of LAI-VPD+1 (green box) and SM-VPD+1 (brown box) during 1982–2018 across the aridity gradient. c Coefficient of LAI-VPD+1 during 1982–1999 (light green box) and 2000–2018 (dark green box) across the aridity gradient. d Coefficient of SM-VPD+1 during 1982–1999 (light brown box) and 2000–2018 (dark brown box) across the aridity gradient. In (a), * indicates significant change trends. In (b–d), LAI-VPD+1 indicates the vegetation path of land‒atmosphere coupling, representing the strength of vegetation‒VPD coupling; SM-VPD+1 indicates the soil moisture path of land‒atmosphere coupling, representing the strength of soil moisture‒VPD coupling. The error bars are the standard deviations of the four SEMs. The percentage values are the medians across the four SEMs. “MMG” (red circle) represents the temperature, VPD, latent heat flux (LH), and sensible heat flux (SH) data derived from the Modem Retrospective Analysis for Research and Applications (MERRA-2), and soil moisture data derived from the Global Land Evaporation Amsterdam Model (GLEAM); “MEG” (blue circle) represents the temperature and VPD data derived from MERRA-2, SH and LH data derived from the European Centre for Medium-Range Weather Forecasts (ERA5-land), and soil moisture data derived from GLEAM; “EEG” (green circle) represents the temperature, VPD, SH and LH data derived from ERA5-land, and soil moisture data derived from GLEAM; and “EMG” (purple circle) represents the temperature and VPD data derived from ERA5-land, SH and LH data derived from MERRA-2, and soil moisture data derived from GLEAM.

Full size image

Vegetation‒VPD coupling contributes 10.19% and 12.36% to the VPD variations in areas where 0.3 < AI ≤ 0.4 and 0.4 < AI ≤ 0.5, respectively (Fig. 1b). The contribution of the VPD variations from soil moisture‒VPD coupling is nearly twice as strong as that of vegetation‒VPD coupling. Specifically, it accounts for 22.92% and 25.18% of the VPD variations in areas where 0.4 < AI ≤ 0.5 and 0.5 < AI ≤ 0.65, respectively (Fig. 1b).

Weak land‒atmosphere coupling is observed in areas with lower vegetation cover and drier soil (AI ≤ 0.3). In these areas, vegetation‒VPD coupling is almost negligible (the average coefficient of LAI-VPD+1 is −0.03, p < 0.05; Fig. 1b), and soil moisture‒VPD coupling is also weaker compared to other areas (the average coefficient of SM-VPD+1 is −0.23, p < 0.05; Fig. 1b).

The impact of ecological restoration on land‒atmosphere coupling

Whether ecological restoration can alleviate atmospheric drought depends on two aspects, one is the change in vegetation/soil moisture, the other is the change in land‒atmosphere coupling.

Both vegetation and soil moisture pathways are intensified after ecological restoration, showing varied impacts on VPD across aridity gradients. In areas where 0.3 < AI ≤ 0.4, ecological restoration with greening and wetting soil mitigates the increase in VPD through the intensified land‒atmosphere coupling. The coefficient of LAI-VPD+1 before and after ecological restoration (1982–1999 and 2000–2018) are 0.10 and −0.31, respectively (p < 0.05; Fig. 1c); the coefficient of SM-VPD+1 before and after ecological restoration are −0.23 and −0.24, respectively (p < 0.05; Fig. 1d). In areas where 0.4 < AI ≤ 0.5, ecological restoration with greening and drying soil mitigates the increase in VPD through the strengthened vegetation‒VPD coupling and the weakened soil moisture‒VPD coupling. In detail, the coefficient of LAI-VPD+1 before and after ecological restoration are 0.00 and −0.21, respectively (p < 0.05; Fig. 1c); the coefficient of SM-VPD+1 before and after ecological restoration are −0.41 and −0.35, respectively (p < 0.05; Fig. 1d). In areas where 0.5 < AI ≤ 0.65, ecological restoration with greening and drying soil amplifies the increase in VPD through the weakened vegetation‒VPD coupling and the intensified soil moisture‒VPD coupling. In detail, the LAI-VPD+1 becomes nonsignificant in the late period of ecological restoration, and the coefficient of SM-VPD+1 before and after ecological restoration are −0.35 and −0.50, respectively (p < 0.05; Fig. 1d).

During the late period of ecological restoration, the contribution of vegetation‒VPD coupling to the VPD variations increases, with its contribution increasing by 8.90% (0.2 < AI ≤ 0.3), 6.09% (0.3 < AI ≤ 0.4), and 12.95% (0.4 < AI ≤ 0.5), but decreasing by 6.10% in areas where 0.5 < AI ≤ 0.65 (Fig. 1c). Interestingly, the increased contribution of soil moisture‒VPD coupling to the VPD variations is not observed in areas with stronger soil moisture‒VPD coupling, but rather in areas where this coupling is relatively weak. In areas where AI > 0.3, ecological restoration enhances soil moisture‒VPD coupling, but the average contribution increases by less than 5%. Conversely, in areas where AI ≤ 0.3, although ecological restoration enhances soil moisture‒VPD coupling to a lesser extent, it amplifies the contribution of this coupling to the VPD variations. Specifically, the contribution of soil moisture‒VPD coupling increases by 21.16% (AI ≤ 0.05), 19.88% (0.05 < AI ≤ 0.2) and 37.51% (0.2 < AI ≤ 0.3), respectively, while it increases by only 0.08% (0.3 < AI ≤ 0.4), 2.86% (0.4 < AI ≤ 0.5), and 8.83% (0.5 < AI ≤ 0.65), respectively (Fig. 1d).

Mechanism of ecological restoration influencing land‒atmosphere coupling

The LightGBM (GBM) models here are applied to attribute vegetation‒VPD coupling and soil moisture‒VPD coupling, to relative humidity (RH), net radiation (Rn), convective available potential energy (CAPE), vertically integrated moisture divergence (VIMD), LH and SH (“Methods”). Changes in energy and moisture fluxes are observed in areas where ecological restoration contributes to land‒atmosphere coupling. In areas where land‒atmosphere coupling is intense (0.3 < AI ≤ 0.65), RH, Rn, and CAPE are the most important variables influencing land‒atmosphere coupling (Figs. 2, 3). After ecological restoration, the strength of land‒atmosphere coupling has been changed due to variations in moisture and energy exchange (Fig. 4).

Fig. 2: The importance of water-heat variables for land‒atmosphere coupling.
figure 2

a, b The importance of water-heat variables for land‒atmosphere coupling in areas where 0.3 < AI ≤ 0.4; vegetation‒VPD coupling (a) and soil moisture‒VPD coupling (b). c, d Same as (a, b), but in areas where 0.4 < AI ≤ 0.5; vegetation‒VPD coupling (c) and soil moisture‒VPD coupling (d). e–f Same as (a, b), but in areas where 0.5 < AI ≤ 0.65; vegetation‒VPD coupling (e) and soil moisture‒VPD coupling (f). The percentage is the importance of the predictor variable for land‒atmosphere coupling. Here variables are latent heat flux (LH), sensible heat flux (SH), convective available potential energy (CAPE), vertically integrated moisture divergence (VIMD), relative humidity (RH), and net radiation (Rn).

Full size image
Fig. 3: The three most important variables for land‒atmosphere coupling.
figure 3

a–i The importance of variables for vegetation‒VPD coupling in areas where 0.3 < AI ≤ 0.65. j–r Same as (a–i), but for soil moisture‒VPD coupling in areas where 0.3 < AI ≤ 0.65. Here variables are latent heat flux (LH), sensible heat flux (SH), convective available potential energy (CAPE), vertically integrated moisture divergence (VIMD), relative humidity (RH), and net radiation (Rn).

Full size image
Fig. 4: Mechanism of ecological restoration influencing land‒atmosphere coupling.
figure 4

Vegetation‒VPD coupling and soil moisture‒VPD coupling are generally negative, representing that increases in LAI or soil moisture would decrease VPD. The “Enhanced” indicates the negative vegetation‒VPD coupling (solid green arrow) or negative soil moisture‒VPD coupling (solid brown arrow) strengthened after ecological restoration. The “Weakened” indicates the negative vegetation‒VPD coupling (hollow green arrow) or negative soil moisture‒VPD coupling (hollow brown arrow) weakened after ecological restoration. (+)/(−) indicates the increase/decrease of variables after ecological restoration. Here variables are latent heat flux (LH), sensible heat flux (SH), convective available potential energy (CAPE), relative humidity (RH), and net radiation (Rn).

Full size image

In areas where 0.3 < AI ≤ 0.4, the most important factors affecting vegetation‒VPD coupling are Rn and CAPE. When CAPE is in the range of 75–130 J kg−1, it is positively correlated with vegetation‒VPD coupling (Fig. 3b). The decrease in CAPE after ecological restoration (from 108.80 to 86.48 J kg−1, Supplementary Fig. 3e) strengthens the negative of vegetation‒VPD coupling. The most important factor for soil moisture‒VPD coupling is RH. It shows a positive correlation with the coupling, and the decreased RH (from 56.80% to 53.35%, Supplementary Fig. 3d) amplifies the negative of soil moisture‒VPD coupling (Fig. 3j). In areas where 0.4 < AI ≤ 0.5, Rn is the most vital factor for vegetation‒VPD coupling, when Rn is about 188 W m−2, the negative of this coupling is the strongest (Fig. 3d). However, the increase in Rn after ecological restoration (from 184.19 to 187.94 W m−2, Supplementary Fig. 3c) slightly weakens the negative of soil moisture‒VPD coupling (Fig. 3n). In areas where 0.5 < AI ≤ 0.65, CAPE and LH are the most important variables for vegetation‒VPD coupling. Although the decrease in LH (from 82.85 to 77.85 W m−2, Supplementary Fig. 3a) enhances the negative of the coupling (Fig. 3h), the reduction in CAPE after ecological restoration (from 259.47 to 213.47 J kg−1) mitigates this negative effect (Fig. 3g). CAPE and SH are crucial for soil moisture‒VPD coupling. Although the reduction in CAPE weakens the negative of the coupling (Fig. 3p), the SH increases after ecological restoration (from 24.35 to 29.85 W m−2; Supplementary Fig. 3b) strengthens this negative effect (Fig. 3q).

Discussion

Our results reveal that ecological restoration in China’s drylands has intensified land‒atmosphere coupling, and we provide a comprehensive understanding of how ecological restoration mitigates atmospheric drought via land‒atmosphere coupling by separating the pathways of vegetation and soil.

Robust land‒atmosphere coupling in drylands is primarily concentrated in transitional areas between dry and wet climates. In arid areas, the absence of vegetation and limited soil moisture restrict the evapotranspiration process, whereas in humid areas, radiation is the limiting factor for evapotranspiration48. Consequently, the increased evapotranspiration due to ecosystem changes is not sufficient to trigger strong land‒atmosphere coupling in these areas. Transitional zones between dry and wet climates are covered by diverse vegetation types, increasing surface roughness and promoting interactions between the land surface and the atmosphere. Due to the active energy and moisture interactions, these areas exhibit intense land‒atmosphere coupling10,49.

A few potential caveats should be considered when interpreting our results. Our analysis relies on the median values derived from the four SEMs. All four SEMs consistently pinpoint regions characterized by intense land‒atmosphere coupling (Fig. 5a–h). In addition, the impact of ecological restoration on land‒atmosphere coupling derived from the four SEMs remains consistent across most areas (Fig. 5i–p). However, for instance, in areas where 0.5 < AI ≤ 0.65, the strength of soil moisture‒VPD coupling before and after ecological restoration shows obvious differences across diverse data sources: in MEG and EEG, ecological restoration mitigates the increase in VPD by weakening soil moisture‒VPD coupling; while in MMG and EMG, ecological restoration amplifies the increase in VPD through enhancing soil moisture‒VPD coupling (Fig. 5m–p). It brings uncertainty into our study, but also prompts us to pay more attention to the impact of data sources on study’s results.

Fig. 5: Coefficients of LAI-VPD+1 and SM-VPD+1 derived from the four SEMs.
figure 5

a–d Coefficients of LAI-VPD+1 derived from the four SEMs: from MMG (a), from MEG (b), from EEG (c), and from EMG (d), respectively. e–h Coefficients of SM-VPD+1 derived from the four SEMs: from MMG (e), from MEG (f), from EEG (g), and from EMG (h), respectively. i–l Coefficients of LAI-VPD+1 during 1982–1999 (orange column) and 2000–2018 (green column) derived from the four SEMs: from MMG (i), from MEG (j), from EEG (k), and from EMG (l), respectively. m–p Coefficients of SM-VPD+1 during 1982–1999 (orange column) and 2000–2018 (green column) derived from the four SEMs: from MMG (m), from MEG (n), from EEG (o), and from EMG (p), respectively.

Full size image

A higher VPD is related to reduced gross primary productivity (GPP), decreased tree height, and lower agricultural yields, among other adverse effects, within drylands50,51. Our study shows that large-scale ecological restoration in China’s drylands mitigates the increase in VPD in semiarid areas (0.3 < AI ≤ 0.5; Fig. 1c, d). That is, ecological restoration has the potential to alleviate atmospheric drought, and prevent further land degradation in these drought areas. However, forest protection and restoration in dry-subhumid areas3 (0.5 < AI ≤ 0.65) might amplify the increase in VPD, aggravating the atmospheric drought. The results emphasize the importance of regional climate background, and tell us that a successful ecological restoration needs to be adapted to local climate conditions. In addition, soil conditions and social development are key factors for the sustainable development of ecological restoration52. We suggest that in areas with serious soil erosion and poor soil water storage capacity, restoration should pay attention to planting density and species selection to prevent aggravating soil and atmospheric drought.

Large-scale restoration in China’s drylands mitigates drought risk from the perspective of land surface water resources53,54. Our results also reveal that ecological restoration mitigates atmospheric drought through land‒atmosphere coupling, which is crucial for fighting extreme climate events in drylands. Extremes not only lead to more ecological droughts55, but also increase the risk of human death17. Heatwaves and reduced precipitation often trigger the onset of droughts in drylands15,56. We further quantify the amplifying effect of climate change (such as warming and precipitation reduction) on atmospheric drought through land‒atmosphere coupling, which is essential for predicting the occurrence probability of extremes. This highlights that ecological restoration mitigates the atmospheric aridification induced by climate change through land‒atmosphere coupling in areas where 0.3 < AI ≤ 0.5, indicating its potential to reduce the probability of concurrent heatwaves and droughts (Fig. 6). Our study emphasizes the importance of land‒atmosphere coupling for extreme climates. In the context of frequent extreme events, this factor cannot be ignored in the study of ecological restoration.

Fig. 6: Impacts of temperature and precipitation changes on VPD through land‒atmosphere coupling.
figure 6

a Coefficients of T-LAI-VPD+1 in 1982–2018 (purple box), 1982–1999 (red box), and 2000–2018 (blue box), respectively. b Coefficients of T-SM-VPD+1 in 1982–2018 (purple box), 1982–1999 (red box), and 2000–2018 (blue box), respectively. c Coefficients of P-LAI-VPD+1 in 1982–2018 (purple box), 1982–1999 (red box), and 2000–2018 (blue box), respectively. d Coefficients of P-SM-VPD+1 in 1982–2018 (purple box), 1982–1999 (red box), and 2000–2018 (blue box), respectively. T-LAI-VPD+1 indicates that temperature affects the next-month VPD through vegetation‒VPD coupling. T-SM-VPD+1 indicates that temperature affects the next-month VPD through soil moisture‒VPD coupling. P-LAI-VPD+1 indicates that precipitation affects the next-month VPD through vegetation‒VPD coupling. P-SM-VPD+1 indicates that precipitation affects the next-month VPD through soil moisture‒VPD coupling. The error bars are the standard deviations of the four SEMs. The percentage values are the medians across the four SEMs. Green circle represents the “MMG” model, orange circle represents the “MEG” model, blue circle represents the “EEG” model, and purple circle represents the “EMG” model.

Full size image

Not all ecological restoration in drylands has the ability to mitigate extremes and promote the sustainable development of ecosystems. Africa’s Great Green Wall (GGW), another important large-scale restoration program, has been implemented to prevent the Sahara Desert from moving south57. In China’s drylands, particularly in areas with similar aridity index (0.05 < AI ≤ 0.2), the government selects grassland as the primary type of ecological restoration. This strategic choice has proven effective in mitigating atmospheric drought and compound extremes, as demonstrated by our analysis (Figs. 1 and 6). On the contrary, in the Sahara region where precipitation is scarce and vegetation largely relies on groundwater, large-scale afforestation is likely to exacerbate soil moisture and groundwater depletion, increasing the probability of extremes58,59. A similar situation is also observed in India, where vegetation greening primarily results from cropland expansion, leading to a less mitigation effect on warming than expected60. Furthermore, it exacerbates groundwater depletion, which is detrimental to ecosystem sustainability. The conclusions drawn from this study cannot directly apply to globally ecological restoration; instead, a comprehensive assessment of restoration on a global scale is necessary.

In conclusion, our study highlights the significance of biogeophysical feedback in influencing climate change. The proposed methods in this study help in understanding the potential of mitigating atmospheric drought through restored ecosystems in other regions. Given the atmospheric circulation, further studies need to focus on the impact of large-scale ecological restoration on global climate change and extreme climate events through land‒atmosphere coupling to promote regional and global sustainable development.

Methods

Study region

Drylands in China are characterized by strong land‒atmosphere coupling due to multilevel transitions in vegetation and soil types61,62,63. The AI within the research region varies across a spectrum, ranging from hyper-arid to humid zones (including hyper-arid (0 < AI ≤ 0.05), arid (0.05 < AI ≤ 0.2), semiarid (0.2 < AI ≤ 0.3, 0.3 < AI ≤ 0.4 and 0.4 < AI ≤ 0.5), dry-subhumid (0.5 < AI ≤ 0.65), and humid (AI > 0.65) subareas, as shown in Fig. 7a).

Fig. 7: Climate and vegetation in northern China during 1982–2018.
figure 7

a The aridity of the study area. b Spatial pattern of the change in LAI. c Spatial pattern of changes in precipitation, VPD and soil moisture. d Spatial pattern of ecological restoration programs. In (b), * indicates a significant change trend of LAI. In (c), PRE, VPD, and SM represent the change trends of precipitation, VPD and soil moisture, respectively. The “+” indicates an increasing trend of hydrological variables and “−” indicates a decreasing trend of hydrological variables. The subareas “i”, “ii”, and “iii” indicate the western Xinjiang, the northwestern Loess Plateau, and the southern Loess Plateau, respectively.

Full size image

The hydrological variables show great changes due to large-scale ecological restoration and climate change. For example, precipitation and soil moisture increase in western areas, such as western Xinjiang (i in Fig. 7c) and the northwestern regions of the Loess Plateau (ii in Fig. 7c), but decrease in northeastern areas of China’s drylands, intensifying water stress in local ecosystems. In southern regions of the Loess Plateau (iii in Fig. 7c), although precipitation increases, but the soil is drier.

Significant afforestation efforts, such as the Three-North Shelter Forest Programme (TNSFP) and the Grain to Green Programme (GTGP), have been implemented in China’s drylands. Large-scale conservation and restoration have led to substantial land improvement and have yielded considerable success in ecosystem recovery, manifesting in significant vegetation greenness (Fig. 7b, d)64. Despite restoration and conservation efforts, climate change has caused human deaths, economic losses, and significant ecosystem damage, such as grain reduction and tree mortality65,66,67,68.

Periods before and after ecological restoration

Notable changes in land cover, evapotranspiration, and soil moisture have been observed in northern China since the GTGP was initiated at the end of 1999. Our study uses regression analysis to identify the trend in LAI changes from 1982 and 2018 and finds an accelerated increase in LAI after 2000 (Supplementary Fig. 4). Therefore, we define the period before 2000 as the stage before ecological restoration (1982–1999), while the period after 2000 is defined as the stage after ecological restoration (2000–2018). These two periods are divided to evaluate the impact of ecological restoration programs on land‒atmosphere coupling.

Lag correlation of land‒atmosphere coupling

We use a 1-month lag correlation methodology to quantify the interactions between the land surface and the atmosphere. This approach considers the potential influence of autocorrelation and potential collinearity between atmospheric variables69,70. In this methodology, partial least squares regression (PLSR) is employed to quantify the strength of land‒atmosphere coupling. As an illustration, the equation that determines the strength of land‒atmosphere coupling is presented here:

$${{{{{{\rm{VPD}}}}}}}_{t+1}={{{{{\rm{a}}}}}}+{{{{{\rm{b}}}}}}\times {{{{{{\rm{Land}}}}}}}_{t}+{{{{{\rm{c}}}}}}\times {{{{{{\rm{VPD}}}}}}}_{t}$$
(1)

The seasonal cycles and long-term trends in all variables are removed, and “t + 1” indicates the time step of 1 month. “Landt” and “VPDt” are the values of the land surface parameters and VPD, respectively, in month “t”. “VPDt+1” denotes the value of the VPD at “t + 1” month. In this equation, “b × Landt” and “c × VPDt” indicate the effects of the land surface (including vegetation and soil) and VPD autocorrelation on VPD, respectively. “b” and “c” indicate the strength of land‒atmosphere coupling and autocorrelation, respectively. Taking the effect of the current-month LAI (LAIt) on the next-month VPD (VPDt+1) as an example, a “b” value of 0.1 means that the next-month VPD increases by 10% of its standard deviation when the current-month LAI increases by one standard deviation. That is, when “b” is greater than 0, the increase in the LAI promotes the increase in the VPD; otherwise, the increase in the VPD is mitigated.

Paths of land‒atmosphere coupling

We first select parameters to construct a network representing interactions between the land surface and the atmosphere, and given the climate autocorrelations, we integrate climate change into the network. Temperature, precipitation, and VPD are the three meteorological factors selected in this study, as they are important factors that affect vegetation growth in drylands. Leaves are important for gas and energy exchange between plants and the atmosphere; therefore, LAI is selected as the indicator of vegetation growth. In drylands, vegetation growth heavily relies on root soil moisture. Given the continuous relationship between surface soil moisture and root soil moisture, the latter is chosen as the indicator of the overall status of soil moisture. Changes in the land surface affect the transfer of moisture and energy back to atmosphere71. Two crucial components, LH and SH, are recognized as major linkages between the land surface and the atmosphere.

This study uses SEM to separate paths of land‒atmosphere coupling. SEM not only encompasses interactions among climate variables but also considers the influence of the meteorological state on the subsequent meteorological state. Specifically, temperature, precipitation, and VPD in a particular month (TEM, PRE, and VPD in SEM) are assumed to impact the corresponding variables in the subsequent month (TEM+1, PRE+1, and VPD+1 in SEM). Based on the lag correlation of land‒atmosphere coupling (Eq. 1), we assume the current-month LAI and soil moisture (LAI and SM in SEM) have impacts on the meteorological variables in the subsequent month. In SEM, partial least squares (PLS) is used to calculate the path coefficients, which excludes the interference of other factors. Therefore, the path coefficients between LAI/SM and VPD+1 (LAI-VPD+1 and SM-VPD+1) are used to indicate the strength of vegetation‒VPD coupling and soil moisture‒VPD coupling, respectively. In addition, considering the lagged effects of climate change on vegetation and soil, the LAI, soil moisture, SH, and LH in SEM are computed as the averages of the current month and the next month (Supplementary Fig. 5).

Energy and moisture fluxes in response to land‒atmosphere coupling

GBM models based on Bayesian optimization are constructed to attribute land‒atmosphere coupling to water-heat variables within each aridity subarea. In this study, SHapley Additive exPlanations (SHAP) values are used to analyze the output of the GBM model. Higher SHAP values for predictor variables indicate their greater importance in predicting parameters. A SHAP value exceeding 0 indicates that the predictor variable positively influences the predicted variable, while a SHAP value falling below 0 indicates a negative influence.

Each GBM model includes six predictor factors. LH and SH are introduced to capture the effects of energy exchange on land‒atmosphere coupling, and Rn can influence the energy exchange between the land surface and the atmosphere. CAPE represents the convection capacity, RH indicates the atmospheric water vapor content, and VIMD indicates the vertical direction of water vapor diffusion. This study aims to test the hypothesis that these aforementioned variables can explain the influence of ecological restoration on land‒atmosphere coupling, referencing previous studies that have highlighted their importance15. GBM and SHAP are applied to estimate the importance of each explanatory variable. The GBM models perform well in capturing land‒atmosphere coupling, with R2 values ranging from 0.71 to 0.89, and are therefore selected for the analysis in this study.

Data selection

The uncertainty of land‒atmosphere coupling comes from methods, data sources, and research scales. To avoid errors introduced by the calculation process, our study selects only the reanalyzed products that do not require further processing. In addition, to reduce potential errors arising from data source selection, we further assess the correlations between different datasets and their change trends to select the datasets used in our study (Supplementary Note 2). Considering the change trends of different datasets and their correlations, the following selection rules are applied in this study: if a variable has a dataset with weak and nonsignificant correlation with others, and its change trend is opposite to that of other datasets, it is excluded from our study. When two datasets exhibit a significant correlation and share the same change trends, we prioritize the dataset with the stronger correlation. If there is a significant correlation between different datasets and have different change trends, we select all datasets. After selecting datasets, monthly precipitation data sourced from the Climatic Research Unit (CRU), monthly temperature, VPD, SH, and LH data retrieved from MERRA-2 and ERA5-land, in addition to soil moisture data at root depth acquired from GLEAM, are selected.

Based on the selected datasets, this study constructs four different SEMs with the following parameters (Supplementary Table 9): Temperature, VPD, LH, and SH from the MERRA-2 dataset (MMG); temperature and VPD from the MERRA-2 dataset, and LH and SH from the ERA5-land dataset (MEG); temperature, VPD, LH, and SH from the ERA5-land dataset (EEG); temperature and VPD from the ERA-land dataset, and LH and SH from the MERRA-2 dataset (EMG). The path coefficients used for analysis need to be verified twice. First, SEM must pass the model verification. The normed fit index (NFI) and standardized root mean square residual (SRMR) are selected as evaluation metrics for assessing the fitness and robustness of SEM. Specifically, SEM is considered valid when the NFI is greater than 0.90 and SRMR is less than 0.0872. We evaluate 84 SEMs encompassing diverse aridity regions, study periods and data sources, and it is noted that all models demonstrate an NFI exceeding 0.9 and an SRMR below 0.08 (Supplementary Table 10). Second, only when the path coefficients are statistically significant (p < 0.05) can they be extracted for analysis. After verification, the median values of the four SEMs are used for subsequent analysis.