Spatiotemporal Responses of Global Vegetation Growth to Terrestrial Water Storage

Abstract

Global vegetation growth is dynamically influenced and regulated by hydrological processes. Understanding vegetation responses to terrestrial water storage (TWS) dynamics is crucial for predicting ecosystem resilience and guiding water resource management under climate change. This study investigated global vegetation responses to a terrestrial water storage anomaly (TWSA) using NDVI and TWSA datasets from January 2004 to December 2023. We proposed a Pearson-ACF time lag analysis method that combined dynamic windowing and enhanced accuracy to capture spatial correlations and temporal lag effects in vegetation responses to TWS changes. The results showed the following: (1) Positive NDVI-TWSA correlations were prominent in low-latitude tropical regions, whereas negative responses occurred mainly north of 30°N and in South American rainforest, covering 38.96% of the global vegetated land. (2) Response patterns varied by vegetation type: shrubland, grassland, and cropland exhibited short lags (1–4 months), while tree cover, herbaceous wetland, mangroves, and moss and lichen typically presented delayed responses (8–9 months). (3) Significant bidirectional Granger causality was identified in 16.39% of vegetated regions, mainly in eastern Asia, central North America, and central South America. These findings underscored the vital role of vegetation in the global water cycle, providing support for vegetation prediction, water resource planning, and adaptive water management in water-scarce regions.

1. Introduction

Covering approximately 77% of global land, vegetation plays a critical role in facilitating energy exchange within the natural environment, serving as a primary dietary component for both humans and animals and contributing to the maintenance of ecological balance [1,2]. The spatial distributions and growth trends in vegetation are dictated by water resources to a great extent [3,4]. Meanwhile, water availability affects more than 50% of the primary productivity within terrestrial ecosystems and influences global carbon cycle through the regulation of evapotranspiration [5,6,7]. Thus, identifying how vegetation growth responds to water availability is vital for understanding global water- vegetation–carbon cycles, alongside facilitating the achievement of the UN Sustainable Development Goals, especially SDG 6.6 [8].

Various hydrological factors, including precipitation, soil moisture, and evapotranspiration, were commonly employed to explore the responses of vegetation growth to water conditions [9,10,11,12,13]. However, these hydrological indicators have certain limitations. Precipitation indirectly provides information about the availability of surface water, soil moisture can be restricted by the measurement accuracy in the root zone [14,15], and evapotranspiration is only the external manifestation of water utilization by vegetation [16]. Therefore, terrestrial water storage (TWS) data have been introduced as a comprehensive indicator for investigating the hydrological effects on vegetation growth in recent years [17,18]. TWS, comprising groundwater, plant canopy water, and soil moisture, reflects all types of water within the terrestrial [19]. TWS has been proven to be effective in capturing seasonal fluctuations and periodic variations in vegetation growth [12,20]. Usually, the TWS anomaly (TWSA) data are obtained from equivalent water thickness data from the Gravity Recovery and Climate Experiment (GRACE) satellite and its follow-on mission GRACE-Follow-On (GRACE-FO). The missing data within the GRACE/GRACE-FO mission and the subsequent 11-month gap can be filled using the Singular Spectrum Analysis (SSA) interpolation method, generating continuous TWSA monthly product at a spatial resolution of 0.25 degrees from January 2004 to December 2023 [21]. This dataset, containing long time series and up-to-date information, can be utilized to detect the latest trends in global TWSAs and then explore the patterns of NDVI responses to TWSAs over two decades.

Vegetation growth is a gradual process that requires the accumulation of sufficient materials and water. Therefore, responses of vegetation to various meteorological and hydrological factors often exhibit certain lag effects [22,23,24]. In recent years, scholars have mainly used two methods to explore the lag effects of vegetation growth in response to TWS: machine learning methods and spatiotemporal statistical methods. Machine learning techniques, such as support vector machines and decision trees, focus on building predictive models to capture relationships between variables and forecast future trends. These methods, while capable of handling complex and non-linear relationships, often come with high computational demands and are prone to overfitting, especially at the pixel level, making them less applicable on large spatial scales [25,26]. In contrast, many commonly used spatiotemporal statistical methods rely on spatial analysis techniques like Pearson correlation analysis, windowed cross-correlation, and geographically weighted regression to estimate time lags by identifying the highest correlation coefficient between vegetation indices and hydrological factors [27]. However, these previous approaches face notable limitations. They often require manual configuration of analysis windows, which can introduce bias and demand long time series to capture sufficient variability. Additionally, they tend to have lower accuracy in estimating time lags over extended periods [27,28]. To address these limitations, our study introduced the Pearson-ACF time lag analysis method, which combines Pearson correlation analysis with the autocorrelation function (ACF). By leveraging the ACF to assess correlations over continuous time series, this method minimizes the need for predefined analysis windows and improves the accuracy of time lag estimations, making it particularly well suited for long-term, global-scale studies.

In this study, using data from the GRACE/GRACE-FO missions and the NDVI from January 2004 to December 2023, we introduced a robust framework for analyzing the responses of global vegetation growth to terrestrial water storage. The specific objectives of this study are the following: (1) to identify global trends in the TWSA and NDVI over the study period; (2) to analyze the spatial heterogeneity of responses and time lags of the NDVI to TWSA for different vegetation types; (3) to explore the causality between the TWSA and NDVI across different regions; (4) to examine the responses of the NDVI to multiple hydrological factors.

2. Data and Methods

2.1. Data

2.1.1. GRACE Data

The global TWSA data were obtained from GRACE/GRACE-FO RL06.2 Mascon Solutions produced by the Center for Space Research (CSR), University of Texas, USA (https://www2.csr.utexas.edu/grace/) (accessed on 3 March 2024) [29]. Compared to the previous RL06 version, CSR RL06.2 mascon grids have undergone corrections tailored for representation on an ellipsoidal earth, which can prevent any signal leakage from land corrections into the ocean [29,30]. The data collection was available from April 2002 to December 2023 at a monthly 0.25° grid resolution, which was calculated relative to the mean baseline from 2004 to 2009. The missing data were effectively addressed using the SSA method to ensure continuity and reliability [21]. In this study, the terrestrial water storage change (TWSC) was calculated as the difference between two adjacent months’ TWSA values to reflect the change in water storage.

2.1.2. NDVI Data

The NDVI, a crucial indicator for gauging vegetation growth, was collected from The Terra Moderate Resolution Imaging Spectroradiometer (MODIS) Vegetation Indices (MOD13Q1 V6.1) (https://lpdaac.usgs.gov/) (accessed on 3 March 2024) [31,32,33]. The MODIS NDVI products have undergone masking for clouds, haze, and heavy aerosols [34]. The dataset spanned from February 2000 to December 2023, with a resolution of 250 m and 16-day intervals, which was further synthesized into monthly products utilizing the Maximum Value Composite (MVC) method to remove the influence of cloud shadows and enhance seasonal signals [35].

2.1.3. Meteorological and Hydrological Data

The meteorological and hydrological data, including precipitation (PRE), evaporation (EVA), and runoff (RUN) data, were derived from the ERA5-Land monthly dataset with a spatial resolution of 0.1° (https://www.copernicus.eu/) (accessed on 10 March 2024). Despite inherent uncertainties, ERA5-Land data have been extensively validated, demonstrating strong reliability across diverse regions and climatic conditions [36,37,38]. The hydrological variables, including plant canopy surface water (PCSW), groundwater storage (GWS), profile soil moisture (PSM), root zone soil moisture (RZSM), and surface soil moisture (SSM), were sourced from the National Aeronautics and Space Administration (NASA) Global Land Data Assimilation System Version 2.2 (GLDAS-2.2) (https://disc.gsfc.nasa.gov/) (accessed on 10 March 2024). Specifically, SSM reflects rapid hydrological responses in the topsoil layer, RZSM represents the water accessible to vegetation in the root zone, and PSM integrates both to capture the full-depth moisture dynamics. These layered soil moisture metrics jointly inform plant water availability and vegetation response at varying temporal scales. Improving on the earlier versions, GLDAS-2.2 employed data assimilation (DA) techniques, enhancing the precision and dependability of the products [39,40]. The daily dataset was available from 2003 to 2023, featuring a spatial resolution of 0.25° grid.

2.1.4. Land Cover Data

The European Space Agency (ESA) WorldCover product was utilized to determine global land cover patterns, featuring a spatial resolution of 10 m (https://viewer.esa-worldcover.org/worldcover) (accessed on 20 March 2024). This map was derived from Sentinel satellite data, renowned for their high-quality imagery and extensive coverage [2]. The WorldCover product categorized global vegetation into 7 distinct classifications, including common types such as tree cover, shrubland, grassland, and cropland and relatively rare types such as herbaceous wetland, mangroves, and moss and lichen, which were profoundly susceptible to water availability. Therefore, the use of this land cover categorization yielded invaluable insights into regions particularly impacted by hydrological dynamics.

Herein, all datasets were processed to calculate monthly anomalies relative to the mean baseline from 2004 to 2009 to remove the seasonal periodicity and resampled into a grid of 0.25° × 0.25° using the nearest neighbor resampling method to align with the TWSA data. Furthermore, our analysis focused exclusively on the period from January 2004 to December 2023, owing to the accessibility and currency of the data.

2.2. Methods

The overall framework of this study is presented in Figure 1: (1) data collection and preprocessing: collecting CSR GRACE, MODIS NDVI, ERA5-Land, and GLDAS-2.2 and conducting preprocessing operations; (2) responses of the NDVI to TWSA: detecting the trends, spatial responses, time lags, and causality between the TWSA and NDVI; (3) responses of the NDVI to multiple hydrological factors, including TWSA, TWSC, PRE, EVA, RUN, GWS, PCSW, SSM, RZSM, and PSM: exploring the relationships between the NDVI and hydrological factors and relationships among hydrological factors. The details are elucidated below.

2.2.1. Singular Spectrum Analysis Gap-Filling

The scattered missing or anomalous data within GRACE (17 months), GRACE-FO (2 months), and the 11-month data gap between the two missions greatly inhibited the comprehensive analysis [41,42]. Here, a data-adaptive approach composed of SSA and cross-validation was utilized to reconstruct the missing data [43,44]. The technique has been demonstrated to be skilled in deducing missing data from fluctuating patterns identified within existing long-term observations [21,45,46].

Given a uniformly sampled time series X = [x₁, x₂, …, x_n], a trajectory matrix Y can be built: Y = [X₁, X₂, X₃, …, X_K], where K = N − M + 1.

Then, the singular value decomposition (SVD) is applied to factorize Y directly:

$${\mathrm{Y}}_{\mathrm{M}\mathrm{K}}={\mathrm{U}}_{\mathrm{M}\mathrm{M}}{\mathrm{E}}_{\mathrm{M}\mathrm{M}}{\mathrm{V}}_{\mathrm{M}\mathrm{K}}^{\prime }$$

(1)

where U, E, and V are the eigenvector, the diagonal matrix, and the orthonormal eigenvector of the lag-covariance matrix, respectively. The Y matrix can be reconstructed through the sum of Z_i, where each mode results from the product of Empirical Orthogonal Functions (EOFs) and principal components (PCs) output by the principal component analysis (PCA) procedure [47]:

$$\mathrm{Y}=\sum _{\mathrm{i}=1}^{\mathrm{L}}{\mathrm{Z}}_{\mathrm{i}}=\sum _{\mathrm{i}=1}^{\mathrm{L}}{\mathrm{E}\mathrm{O}\mathrm{F}}_{\mathrm{i}}\times {\mathrm{P}\mathrm{C}}_{\mathrm{i}}^{\prime }=\sum _{\mathrm{i}=1}^{\mathrm{L}}\left({\mathrm{U}}_{*,\mathrm{i}}{\mathrm{\sigma }}_{\mathrm{i}}\right)\times \left({\mathrm{V}}_{*,\mathrm{i}}\right)$$

(2)

where $\mathrm{L}=\mathrm{m}\mathrm{i}\mathrm{n}(\mathrm{M},\mathrm{K})$. Then, the skew-diagonal element is used to generate reconstructed components (RCs). Accordingly, the sum of all RCs equals the filled original time series X:

$$\mathrm{X}=\sum _{\mathrm{i}=1}^{\mathrm{L}}{\mathrm{R}\mathrm{C}\mathrm{s}}_{\mathrm{i}}^{\mathrm{p}}=\sum _{\mathrm{i}=1}^{\mathrm{L}}\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\left({\mathrm{Z}}_{\mathrm{i}}^{\mathrm{j},\mathrm{k}}\right)$$

(3)

where only Z_i satisfying j + k = p +1 (p = 1, 2, …, N) are used.

In our study, the iterative strategy was utilized to gradually update the gap value and ensure the reliability of data [45]. Two loops were added to control the complexity of the reconstruction process, which separately relied on the window width (M) and the number of RCs (L) [48].

Hence, when the SSA-filling method was applied for the missing data within the GRACE or GRACE-FO mission (SSA-filling-a), the optimal values of M and L were uniformly determined in advance. For the 11-month gap between GRACE and GRACE-FO (SSA-filling-b), the parameters were fine-tuned via cross-validation to achieve optimal reconstruction. Additionally, the error and uncertainty of the filled data were assessed through fitting residuals.

2.2.2. Seasonal Signal Removal and Mann–Kendall Trend Analysis

To remove the strong seasonality present in the NDVI and other hydroclimatic variables, we calculated monthly anomalies relative to a multi-year climatological baseline (2004–2009). This approach enhances the detection of interannual variability and ensures consistency across datasets. The chosen baseline aligns with the CSR GRACE TWSA product, which provides values as anomalies relative to the same period. Consequently, the NDVI, precipitation, runoff, and related variables were processed using the uniform deseasonalization method. The anomaly for each variable X is calculated as follows:

$${\mathrm{X}}_{\mathrm{a}\mathrm{n}\mathrm{o}\mathrm{m}\mathrm{a}\mathrm{l}\mathrm{y}}(\mathrm{i},\mathrm{j})=\mathrm{X}(\mathrm{i},\mathrm{j})-{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\mathrm{X}(\mathrm{i},\mathrm{j})$$

(4)

where $\mathrm{X}(\mathrm{i},\mathrm{j})$ represents the raw value for year i and month j, and the second term is the long-term monthly mean over n years.

In this study, the Mann–Kendall (MK) test was applied to detect the presence of a significant monotonic change trend, while the extent of the trend, if detected, was quantified by Sen’s slope [49,50,51,52]. Additionally, a Z-statistic for testing and the p-value of the statistic were estimated. Here, the MK trend test, along with Sen’s slope, was employed to identify tendencies in the TWSA and NDVI from January 2004 to December 2023. Additionally, trends were considered significant at the 95% confidence level if the p-value was less than 0.05.

2.2.3. Pearson-ACF Time Lag Analysis

Pearson correlation analysis is broadly utilized to assess the magnitude and direction of the linear relationship between two continuous variables [53,54,55]. The formula is as follows [56]:

$${\mathrm{\rho }}_{\mathrm{X},\mathrm{Y}}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{X}}_{\mathrm{i}}-\overline{\mathrm{X}}\right)\left({\mathrm{Y}}_{\mathrm{i}}-\overline{\mathrm{Y}}\right)}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{X}}_{\mathrm{i}}-\overline{\mathrm{X}}\right)}^2{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{Y}}_{\mathrm{i}}-\overline{\mathrm{Y}}\right)}^2}}}$$

(5)

where n is the length of the time series; ${\mathrm{X}}_{\mathrm{i}}$ is the value of variable $\mathrm{X}$ at time i; ${\mathrm{Y}}_{\mathrm{i}}$ is the value of variable $\mathrm{Y}$ at time i; $\overline{\mathrm{X}}$ is the average of the variable $\mathrm{X}$; $\overline{\mathrm{Y}}$ is the average of the variable $\mathrm{Y}$.

Here, the Pearson correlation analysis was implemented to explore the relationships between the TWSA and NDVI during the period of January 2004 to December 2023. After that, the coefficients were categorized into seven classes: strong negative (−1 ≤ ρ < −0.6), medium negative (−0.6 ≤ ρ < −0.4), weak negative (−0.4 ≤ ρ < −0.2), weak positive (0.2 < ρ < 0.4), medium positive (0.4 ≤ ρ < 0.6), strong positive (0.6 ≤ ρ ≤ 1), and non-significant correlation (−0.2 ≤ ρ ≤ 0.2 or p-value ≥ 0.05) to facilitate subsequent analysis.

The autocorrelation function (ACF) is a function used to analyze the autocorrelation degree between time series and itself under different lag orders, which is also essential for detecting patterns, trends, and seasonality [57,58,59]. The formula for computing the ACF at lag k is as follows:

$${\mathrm{A}\mathrm{C}\mathrm{F}}_{\mathrm{k}}={\displaystyle \frac{{\displaystyle {\sum }_{\mathrm{t}=\mathrm{k}+1}^{\mathrm{n}}}\left({\mathrm{x}}_{\mathrm{t}}-\overline{\mathrm{x}}\right)\left({\mathrm{x}}_{\mathrm{t}-\mathrm{k}}-\overline{\mathrm{x}}\right)}{{\displaystyle {\sum }_{\mathrm{t}=1}^{\mathrm{n}}}{\left({\mathrm{x}}_{\mathrm{t}}-\overline{\mathrm{x}}\right)}^2}}$$

(6)

where ${\mathrm{x}}_{\mathrm{t}}$ represents the observation at time t, and $\overline{\mathrm{x}}$ represents the mean. Normally, the horizontal axis of the function depicts the lag period, while the vertical axis represents the degree of correlation. Simultaneously, when the lag order of the ACF curve is 0, the correlation coefficient is 1, suggesting a complete positive correlation with itself. With the increase in the lag order, the ACF curve decreases and gradually approaches zero, indicating that the correlation gradually weakens [60,61].

Vegetation growth is a gradual accumulation process, resulting in a lagged response of the NDVI to various hydrological and meteorological factors [23,62]. Here, a Pearson-ACF method was proposed to estimate the specific time lags of the NDVI to TWSA at a global pixel scale:

(1)

To calculate the autocorrelation function values of the TWSA and NDVI; hence, a functional relationship between autocorrelation values and lag periods was established for each pixel.

(2)

To determine whether the lag conditions were satisfied: (a) the function value of NDVI exceeded half of TWSA for the first time; (b) the function value of NDVI fell within a threshold range (±1.96/$\surd n$, where n represents the sample size), which indicated the results were significant.

(3)

To calculate the Pearson coefficients between the TWSA and NDVI and then judge whether the coefficients were positive or negative. For positive pixels, the time lag was equal to the first lag value that satisfied the conditions; otherwise, the lag was the difference between the NDVI cycle and the first lag value that satisfied the conditions.

(4)

To iterate through each pixel, a global spatial distribution map of time lags for NDVI responses to TWSA changes was obtained.

By considering the lagged effects between the NDVI and TWSA and the precise characterization of pixel-level positive and negative responses, the time lags were accurately calculated. The processing flow, characterized by its simplicity, efficiency, accuracy, and universality, was utilized to dynamically determine the time lags and advance understanding of the relationships between the NDVI and TWSA at a large-scale and long-term pixel level. Through empirical validation, this modified method overcame the limitations of traditional approaches, such as the manual setting of analysis windows and inadequate time series or windows [27]. The time lag estimation method applied here is of paramount importance in elucidating the spatiotemporal associations between vegetation growth and TWS, providing robust support for future predictions of carbon and water cycles.

2.2.4. Granger Causality Test

The non-linear Granger causality test was employed to explore the interactions between the TWSA and NDVI, which was completed according to the assumption that if one time series is useful for predicting another time series, then the former is said to be a Granger cause of the latter [63,64,65,66].

The prerequisite for conducting a Granger causality test is that the time series must be stationary. Therefore, the Augmented Dickey–Fuller (ADF) test should be passed first [67,68,69]. Then, the fundamental principle of the Granger test involves two models: one containing only the vector autoregressive (VAR) terms of the independent variable and the other incorporating VAR terms, both the independent and dependent variables [65,70,71]. Specifically, considering two time series ${\mathrm{X}}_{\mathrm{t}}$ and ${\mathrm{Y}}_{\mathrm{t}}$, representing the independent and dependent variables, respectively, the null hypothesis states that ${\mathrm{X}}_{\mathrm{t}}$ has no Granger causal effect on ${\mathrm{Y}}_{\mathrm{t}}$. The two models can be expressed as follows:

$${\mathrm{Y}}_{1\mathrm{t}}={\sum }_{\mathrm{i}=1}^{\mathrm{p}}{\mathrm{\beta }}_{\mathrm{i}}{\mathrm{Y}}_{\mathrm{t}-\mathrm{i}}+{\mathrm{\epsilon }}_{1\mathrm{t}}$$

(7)

$${\mathrm{Y}}_{2\mathrm{t}}={\sum }_{\mathrm{i}=1}^{\mathrm{p}}\left({\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{X}}_{\mathrm{t}-\mathrm{i}}+{\mathrm{\beta }}_{\mathrm{i}}{\mathrm{Y}}_{\mathrm{t}-\mathrm{i}}\right)+{\mathrm{\epsilon }}_{2\mathrm{t}}$$

(8)

where p is the corresponding time lag, ${\mathrm{\alpha }}_{\mathrm{i}}$ and ${\mathrm{\beta }}_{\mathrm{i}}$ are regression coefficients, and ${\mathrm{\epsilon }}_{1\mathrm{t}}$ and ${\mathrm{\epsilon }}_{2\mathrm{t}}$ are the error terms, respectively.

Whether ${\mathrm{X}}_{\mathrm{t}}$ significantly contributed to the prediction of ${\mathrm{Y}}_{\mathrm{t}}$ was evaluated by comparing the residual sum of squares (RSSs) of ${\mathrm{Y}}_{2\mathrm{t}}$ with that of ${\mathrm{Y}}_{1\mathrm{t}}$. If the inclusion of ${\mathrm{X}}_{\mathrm{t}}$ resulted in a significant decline in the RSSs, the null hypothesis was rejected, suggesting that ${\mathrm{X}}_{\mathrm{t}}$ was the Granger cause of ${\mathrm{Y}}_{\mathrm{t}}$ [72,73].

In our study, the Granger causality test was employed to detect hotspots of interactions between vegetation growth and terrestrial water storage for each pixel. If incorporating the TWSA as a predictor improved the NDVI prediction without reciprocal improvement in the TWSA prediction by including the NDVI, the TWSA was deemed the unidirectional Granger cause of the NDVI. Conversely, if the NDVI enhanced the TWSA prediction without reciprocal improvement in the NDVI prediction by including the TWSA, the NDVI was considered the unidirectional Granger cause of the TWSA. Bidirectional causality was inferred if the TWSA improved the NDVI prediction and the NDVI similarly improved the TWSA prediction. While these findings provided insights into potential relationships, they should be interpreted with caution, as Granger causality highlighted predictive linkages rather than direct physical causality. Subsequently, a comprehensive global distribution map illustrating the Granger causality between the NDVI and TWSA was generated.

2.2.5. Regression Analysis

To investigate the correlation between the NDVI and various hydrological factors, basic linear regressions were performed. The regression model can be expressed as follows:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{a}\ast \mathrm{H}\mathrm{y}\mathrm{d}\mathrm{r}\mathrm{o}\mathrm{l}\mathrm{o}\mathrm{g}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{F}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{o}\mathrm{r} + \mathrm{\epsilon }$$

(9)

where a represents the basic linear regression coefficient, and $\mathrm{\epsilon }$ is the intercept. This model was applied separately to each of the ten hydrological factors to estimate their individual contributions to NDVI variations.

Given the potential complexity of interactions between hydrological factors, multiple linear regression analyses were also conducted. These models were constructed to quantify the relationships between vegetation dynamics and various hydrometeorological drivers based on theoretical and physical relevance. Specifically, Equation (10) assesses how the total water availability and surface processes directly influence the NDVI. Equation (11) examines the effect of plant-accessible water and atmospheric water demand on vegetation growth. Equation (12) is conceptually based on the composition of terrestrial water storage (TWS = GWS + PCSW + SSM) and is designed to examine how the GRACE-derived TWSA can be decomposed into interpretable components. Equation (13) follows the classical water balance equation (TWSC = PRE − EVA − RUN) and evaluates the short-term drivers of terrestrial water storage changes. The regression models are expressed as follows:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{a}\ast \mathrm{T}\mathrm{W}\mathrm{S}\mathrm{A}+\mathrm{b}\ast \mathrm{P}\mathrm{R}\mathrm{E}+\mathrm{c}\ast \mathrm{R}\mathrm{U}\mathrm{N}+\mathrm{d}\ast \mathrm{S}\mathrm{S}\mathrm{M} + \mathrm{\epsilon }$$

(10)

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{a}\ast \mathrm{E}\mathrm{V}\mathrm{A}+\mathrm{b}\ast \mathrm{G}\mathrm{W}\mathrm{S}+\mathrm{c}\ast \mathrm{P}\mathrm{C}\mathrm{S}\mathrm{W}+ \mathrm{\epsilon }$$

(11)

$$\mathrm{T}\mathrm{W}\mathrm{S}\mathrm{A}=\mathrm{a}\ast \mathrm{G}\mathrm{W}\mathrm{S}+\mathrm{b}\ast \mathrm{P}\mathrm{C}\mathrm{S}\mathrm{W}+\mathrm{c}\ast \mathrm{S}\mathrm{S}\mathrm{M}+ \mathrm{\epsilon }$$

(12)

$$\mathrm{T}\mathrm{W}\mathrm{S}\mathrm{C}=\mathrm{a}\ast \mathrm{P}\mathrm{R}\mathrm{E}+\mathrm{b}\ast \mathrm{E}\mathrm{V}\mathrm{A}+\mathrm{c}\ast \mathrm{R}\mathrm{U}\mathrm{N}+ \mathrm{\epsilon }$$

(13)

where a-d represent the regression coefficients that were estimated for each model, and $\mathrm{\epsilon }$ is the intercept. To ensure the reliability of the multiple linear regression analyses, it is essential to avoid multicollinearity among factors. We tested for multicollinearity using the Variance Inflation Factor (VIF), with values ranging from 1 to 10, which indicated no multicollinearity.

The primary advantage of multiple linear regression over basic linear regression is the ability to account for the effects of multiple predictors simultaneously. While basic linear regression examines individual factors, multiple linear regression can account for interactions and the collective impact of several variables, thereby improving the robustness for assessing how hydrological factors jointly influence the NDVI.

3. Results

3.1. Trends in the TWSA and NDVI

To explore the temporal characteristics of the TWSA and NDVI, the monthly and yearly averaged global variations are displayed in Figure 2. It was observed that the TWSA showed a prominent decreasing trend and exhibited significant seasonal patterns. Generally, the TWSA reached the peak in April and the valley in September or October each year. This pattern was attributable to lower temperatures during the January–April and October–December periods, culminating in less glacier ablation and evaporation, thereby augmenting the TWSA dominated by precipitation [74]. Conversely, a warm climate from April to October intensified evaporation and transpiration, along with accelerated glacier melting, resulting in a notable TWS reduction [75]. In contrast, the annual NDVI remained stable between 0.30 and 0.35 but exhibited considerable fluctuations and an asymmetrical distribution on a monthly scale, showing a distinct right-skewed pattern. The narrow box widths and low dispersion showed a pronounced seasonality. Additionally, the global average NDVI reached its minimum in February, primarily reflecting winter conditions in the northern hemisphere, where lower temperatures, reduced precipitation, and shorter daylight limit vegetation growth. Conversely, the maximum NDVI occurred in July, driven by the northern hemisphere summer, with higher temperatures, increased precipitation, and longer daylight promoting vigorous plant growth [76]. In the southern hemisphere, these seasonal patterns are reversed, but the larger land areas and more extensive vegetation activity in the northern hemisphere dominate global trends. Furthermore, the 3-month lag in the NDVI response to the TWSA was inferred by comparing the peaks of their respective curves, as shown in Figure 2c,d. The peak in the TWSA occurred around April, while the NDVI reached its maximum in July, indicating this lag.

Trends in the TWSA across continents from January 2004 to December 2023 are shown in

Table 1. Trends in TWSA from 2004 to 2023 and errors of SSA-filling-a/b.

Region	TWSA Trend (cm/Month)	TWSA Trend Error (cm/Month)	R2 of TWSA Trend	SSA-Filling-a Error (cm)	SSA-Filling-b Error (cm)
All Land	−0.0278	0.0041	0.4285	0.2343	0.4492
Asia	−0.0245	0.0027	0.5739	0.4182	0.6373
North America	−0.0764	0.0067	0.6736	0.3494	0.9976
Europe	−0.0620	0.0089	0.4407	0.5703	1.1219
Africa	0.0253	0.0027	0.5835	0.3790	0.9994
South America	−0.0170	0.0113	0.0342	0.6140	1.5299
Oceania	0.0060	0.0052	0.0214	0.6133	1.9823
Greenland	−0.7262	0.0113	0.9850	0.9894	2.5240

, along with standard errors and the coefficient of determination (R2). A significant decline (−0.0278 ± 0.0041 cm/month) was observed in the TWSA of global land (excluding Antarctica) as a whole. The largest decrease occurred in Greenland (−0.7262 ± 0.0113 cm/month), primarily due to rapid glacier and ice-sheet melting caused by global warming. Among the six analyzed continents, a noticeable increase in the TWSA was observed in tropical regions, including Africa (0.0253 ± 0.0027 cm/month) and Oceania (0.0060 ± 0.0052 cm/month). In contrast, significant losses were concentrated in mid-latitudes, particularly in Asia (−0.0245 ± 0.0027 cm/month), North America (excluding Greenland, −0.0764 ± 0.0067 cm/month), Europe (−0.0620 ± 0.0089 cm/month), and South America (−0.0170 ± 0.0113 cm/month). The coefficient of determination (R2) highlighted the reliability of the observed trends. Greenland exhibited the strongest fit (R2 = 0.985), confirming consistent and significant water storage loss. Conversely, Oceania and South America displayed lower R2 values (0.021 and 0.034, respectively), indicating greater variability or noise in their TWSA trends. Errors associated with SSA-filling methods (a and b) are also summarized in Table 1. SSA-filling-a generally showed smaller errors across all regions compared to SSA-filling-b, demonstrating higher reliability for reconstructing missing data in fluctuating TWSA patterns.

To explore the spatial distribution characteristics, the MK trends in the TWSA and NDVI from January 2004 to December 2023 were conducted (Figure 3). Obviously, strong decreases in the TWSA were shown over about half of the global land (49.04%), mainly distributed in Eurasia and North America. Conversely, strong increases were presented over 32.37% of the land, in Africa, eastern North America, and Australia. Moreover, significant vegetation greening was observed in Eurasia, South America, and central Africa, accounting for 39.33% of the global vegetation-covered regions, while pronounced browning was evident in central North America, northern Africa, and western Australia, accounting for 11.14%. Furthermore, consistent trends between the TWSA and NDVI were shown for 19.09% gathered in predominantly arid regions, with 13.62% exhibiting significant increasing trends mainly concentrated in central Africa and northwest China, and 5.47% showing significant decreases in western Australia and northern Africa. Additionally, inconsistent trends between the TWSA and NDVI were exhibited over 19.05% of the vegetation-covered regions.

Furthermore, an in-depth exploration was conducted to investigate the consistency of trends in the TWSA and NDVI corresponding to different vegetation types of global land (Figure 4). The vegetation-covered regions were categorized into seven types according to ESA WorldCover [2]: tree cover (10), shrubland (20), grassland (30), cropland (40), herbaceous wetland (90), mangroves (95), and moss and lichen (100). Here, herbaceous wetlands, mangroves, and moss and lichen, which were relatively uncommon but highly responsive, were considered to comprehensively assess the influence of TWS on vegetation growth. The results revealed that the proportion of areas exhibiting consistent trends notably exceeded those displaying inconsistent trends for shrubland, grassland, cropland, and mangroves, indicating significant positive responses. Conversely, significant negative responses were found for tree cover, herbaceous wetlands, and moss and lichen, as shown in Figure 4b. The largest area ratio with consistent trends was featured for shrubland, while the largest ratio with inconsistent trends was shown for tree cover. This qualitative analysis of the relationships between TWS and vegetation growth across different vegetation types provided a preliminary understanding, which served as the basis for subsequent quantitative analyses, including Pearson correlation and time lag analysis.

3.2. Spatial Responses of the NDVI to TWSA

The spatial responses of vegetation growth to TWS from January 2004 to December 2023 were comprehensively quantified by Pearson correlation analysis at a 95% confidence level (Figure 5). The Pearson coefficients were categorized into seven classes for statistical analysis. The positive responses (0.2 < ρ ≤ 1) of the NDVI to TWSA were observed to be primarily concentrated in low-latitude tropical regions at a 95% confidence level, covering 38.34% of the global vegetation-covered regions, which were predominantly comprised of shrubland, grassland, and cropland. Conversely, hotspots of negative responses (−1 ≤ ρ < −0.2) were mainly distributed in regions north of 30°N and tropical rainforest areas in South America, with tree cover accounting for 38.96%. The global spatial patterns were roughly consistent with the findings in Figure 3 and Figure 4 in Section 3.1.

Significant strong positive correlations (0.6 ≤ ρ ≤ 1) between the TWSA and NDVI were observed in India, northern Australia, eastern Africa, and eastern South America, covering 7.88% of the global vegetation-covered regions. These regions were characterized by tropical arid conditions with high temperatures, abundant sunlight, and substantial anthropogenic influence, yet they exhibited low rainfall and soil moisture and are regarded as water-limited territories [77,78]. Therefore, increases (decreases) in TWS may enhance (inhibit) vegetation photosynthesis due to limited water availability, leading to a notably positive correlation between vegetation growth and TWS.

On the contrary, significant strong negative correlations (−1 ≤ ρ ≤ −0.6) were primarily exhibited in northeastern North America and northern Europe, accounting for only 2.08%. Meanwhile, these regions were characterized by temperate maritime climates with abundant water storage, which are already suitable for vegetation growth and not limited by water availability [77,78]. Hence, excessive (insufficient) water storage may induce plant water logging (water stress), generating a significant negative correlation between vegetation growth and TWS.

Different types of vegetation exhibit varying responses to terrestrial water storage at different latitudes. To explore the spatial responses of the NDVI to TWSA for different vegetation types, the corresponding proportions pertaining to seven categories of Pearson correlation coefficients were computed (Figure 6). Amidst the seven types, the responses were the most pronounced in shrubland, exhibiting the highest average Pearson coefficient (0.40), while moss and lichen exhibited the lowest average coefficient (−0.33). Meanwhile, mangroves (0.27), cropland (0.25), and grassland (0.13) generally demonstrated significant positive responses, whereas tree cover (−0.11) and herbaceous wetland (−0.20) exhibited distinct negative correlations. Furthermore, weak correlations were predominantly observed, followed by medium and strong correlations, regardless of positive or negative correlations.

3.3. Time Lags of NDVI to TWSA

The growth of vegetation relied on the availability of moisture and nutrients within the soil, which were influenced by various factors such as soil moisture changes, seasonal variations in precipitation, and climatic conditions [79]. Therefore, the responses of vegetation growth to TWS often exhibited apparent lag effects, necessitating consideration of the time lags of NDVI on TWSA changes.

To determine the lags, the autocorrelation function values of TWSA and NDVI for each pixel were calculated. The ACF plots for different vegetation types are shown in Figure 7. The time lags (in months) were depicted along the horizontal axis, while the corresponding correlation values were illustrated along the vertical axis. Notably, it was observed that the function values of TWSA and NDVI typically peaked at a lag period of 12 months, suggesting a consistent 12-month periodicity.

Combining the Pearson correlation results in Figure 5 and Figure 6, the lags of different vegetation types can be inferred based on the proposed Pearson-ACF method. As shown in Figure 7, the lags for shrubland, grassland, cropland, and mangroves were determined as the first time when the NDVI function value fell to less than half of the TWSA function value within the threshold range, corresponding to 3 months, 3 months, 3 months, and 1 month, respectively. The lags for tree cover, herbaceous wetland, and moss and lichen were determined by subtracting the first time satisfying the conditions from the 12-month NDVI cycle, resulting in lags of 9 months, 9 months, and 10 months, respectively.

Based on the Pearson-ACF method, the time lags of NDVI to TWSA at the global pixel scale are shown in Figure 8. It was observed that the lags of vegetation growth to TWS ranged from 0 to 13 months. Shorter time lags (0–4 months) were mainly distributed in arid regions, including Australia, southern Africa, and western North America. Conversely, longer lags (9–13 months) were predominantly situated in moist regions across the northern middle and high latitudes, encompassing areas like southern China, eastern Russia, and Europe.

Histograms of the time lags for various vegetation types are illustrated in Figure 9. Obviously, the responses of shrubland were generally rapid, with lags of mostly 1 month, attributed to the shallow root systems. Grassland and cropland exhibited lags predominantly ranging from 1 to 4 months. In contrast, tree cover, herbaceous wetlands, mangroves, and moss and lichen typically represented longer time lags, around 8–9 months.

3.4. Causality Between the TWSA and NDVI

The effects of TWS on vegetation growth and the reciprocal impacts of vegetation changes on TWS were investigated through the Granger causality test (Figure 10). The TWSA was recognized as the primary unidirectional cause of the NDVI over 19.39% of global vegetation-covered regions, primarily occurring in semi-arid regions such as central North America, central Europe, and central Africa. Vegetation growth in these areas was primarily regulated and impacted by TWS. Conversely, the NDVI was regarded as the Granger unidirectional cause of the TWSA only in 5.98% of vegetation-covered regions, mainly distributed in northern high-latitude regions such as Europe and eastern Asia. Vegetation changes in these regions strongly affected changes in TWS. Additionally, significant bidirectional causality was observed in 16.39% of global vegetation-covered regions, predominantly located in eastern Asia, central North America, and central South America, demonstrating prominent interactions between the TWSA and NDVI.

The influence of TWS on vegetation growth generally outweighed the impact of vegetation growth on TWS and was more prevalent across different vegetation types. Among seven types, tree cover exhibited the highest area ratio of bidirectional causality between the TWSA and NDVI, followed by moss and lichen, cropland, grassland, herbaceous wetland, and shrubland, with mangroves exhibiting the lowest proportion of bidirectional causality. The unidirectional influence of the TWSA on the NDVI was the most significant in tree cover, while the unidirectional influence of the NDVI on the TWSA was the most significant in cropland.

3.5. Responses of the NDVI to Hydrological Factors

Many studies have investigated the responses of vegetation growth to changes in various hydrological factors, such as precipitation, evaporation, and soil moisture [27,80]. Here, we selected 10 variables, including TWSA, TWSC, PRE, EVA, RUN, GWS, PCSW, SSM, RZSM, and PSM, for a comprehensive analysis. These factors were chosen to capture the primary drivers influencing vegetation growth, spanning water balance components and soil moisture conditions. To remove seasonality, all variables were normalized to anomalies relative to the baseline averages from 2004 to 2009, and global averages were calculated. Figure 11a presents the correlations between the NDVI and these factors without accounting for time lags, while Figure 11b incorporates time-lagged correlations to capture delayed responses.

The results revealed significant relationships between the NDVI and hydrological factors, underscoring the complexity of vegetation–hydrology interactions. Positive correlations were observed between the NDVI and PRE, EVA, RUN, and PCSW, reflecting the role of these factors as water inputs that directly support vegetation growth. In contrast, the NDVI exhibited negative correlations with TWSA, TWSC, GWS, SSM, RZSM, and PSM, highlighting the complex interplay between soil moisture and vegetation growth. Among these, the relationship between the NDVI and TWSA was weaker compared to RZSM or SSM. This finding highlighted that vegetation growth is more directly influenced by soil moisture, which serves as an intermediary variable linking the TWSA and NDVI. SSM and RZSM are particularly important, as they directly regulate root-zone and surface-level water availability, which are critical for transpiration and photosynthesis. The negative correlations may also reflect the depletion of water resources in areas of high vegetation activity, where increased growth intensifies water demand.

Figure 11b and

Table 2. Correlations and lags between the NDVI and hydrological factors.

	Correlation Without Considering Lags	Time Lags (Months)	Correlation Considering Lags
TWSA	−0.293	3	−0.707
TWSC	−0.914	0	−0.914
PRE	0.867	0	0.867
EVA	0.955	0	0.955
RUN	0.757	1	0.896
GWS	−0.207	2	−0.519
PCSW	0.799	−1	0.888
SSM	−0.666	1	−0.501
RZSM	−0.516	3	−0.566
PSM	−0.307	3	−0.401

highlight the time-lagged relationships between the NDVI and influencing factors, revealing key patterns in vegetation responses. It can be observed that the lags of the NDVI to TWSC, PRE, and EVA were 0 months, suggesting that vegetation growth promptly reacted to precipitation and evaporation. In other words, precipitation had an immediate impact on vegetation growth, while changes in vegetation growth affected the water cycle through vegetation evaporation and transpiration. Short lags were observed for RUN, SSM, and GWS (1–2 months), reflecting the close connection between vegetation growth and surface or groundwater availability during periods of high water demand. In contrast, longer lags (3 months) were seen for TWSA, RZSM, and PSM, emphasizing the role of sustained hydrological processes, such as root-zone water uptake and deeper soil moisture recharge, in supporting vegetation under water-stressed conditions. Notably, PCSW led NDVI changes by 1 month, underscoring its role as a precursor indicator of vegetation activity. After accounting for time lags, the correlations between the NDVI and hydrological factors strengthened, as shown in Figure 11b. For instance, the correlation with the TWSA strengthened, changing from −0.293 to −0.707, and those with RZSM and GWS also improved significantly. This reinforced the importance of deeper water sources in sustaining vegetation growth.

According to the correlations between variables, five comparative experiments were conducted to study the interactions between vegetation growth and various hydrological indicators, along with the inherent connections among these hydrological factors (

Table 3. Comparative experimental results of the relationships between the NDVI and hydrological factors.

	(1) Coef	(2) Coef	VIF	(3) Coef	VIF	(4) Coef	VIF	(5) Coef	VIF
TWSA	−0.447	−0.150	3.80
TWSC	−1.352
PRE	1.421	0.824	1.97					0.121 (**)	5.35
EVA	1.037			0.872	2.11			−0.841	9.11
RUN	1.012	0.561	2.83					0.134	4.11
GWS	−0.319			−0.210	1.02	1.009	2.62
PCSW	1.151			0.287	2.14	−0.306	1.47
SSM	−1.058	−0.277	3.96			−0.203	3.28
RZSM	−0.810
PSM	−0.485
_cons		0.232		0.220		−0.009		−0.069
R2		0.898		0.955		0.883		0.868

Note: (1) Linear regression coefficients for 10 hydrological factors on NDVI. (2) Multiple linear regression coefficients and VIF for TWSA, PRE, RUN, and SSM on NDVI. (3) Multiple linear regression coefficients and VIF for EVA, GWS, and PCSW on NDVI. (4) Multiple linear regression coefficients and VIF for GWS, PCSW, and SSM on TWSA. (5) Multiple linear regression coefficients and VIF for PRE, EVA, and RUN on TWSC. _cons represented the constant term; R² represented the corresponding fit metric. Only 0.121 (**) was significant at the 95% level (p < 0.05), while the remaining regression coefficients were significant at the 99% level (p < 0.01).

). The basic linear regressions of the NDVI on ten hydrological factors were performed, with regression coefficients maintaining the same polarity as Table 2. Multiple linear regression analyses were conducted for the NDVI and two sets of hydrological factors: (1) TWSA, PRE, RUN, and SSM and (2) EVA, GWS, and PCSW, which exhibited weak correlations and passed multicollinearity tests (the values of the Variance Inflation Factor (VIF) were within the range of 1–10, indicating no multicollinearity). The models demonstrated strong fit with R2 of 0.898 and 0.955, respectively, indicating robustness in the predictive capabilities. (1) The NDVI demonstrated significant positive responses to PRE and EVA, while showing significant negative responses to TWSA and SSM. Specifically, the linear formula for this model was NDVI = −0.150 × TWSA + 0.824 × PRE + 0.561 × RUN − 0.277 × SSM + 0.232. (2) The NDVI showed significant positive responses to EVA and PCSW, while displaying significant negative responses to GWS, namely, NDVI = 0.872 × EVA − 0.210 × GWS + 0.287 × PCSW + 0.220. Then, the relationships between hydrological factors were explored by conducting multiple linear regression analyses for the TWSA and the components (GWS, PCSW, and SSM), yielding a model (TWSA = 1.009 × GWS − 0.306 × PCSW − 0.203 × SSM − 0.009) with an R2 of 0.883. Regression analyses were performed for TWSC and hydrological factors relevant to the water balance equation (PRE, EVA, and RUN), resulting in a model (TWSC = 0.121 × PRE − 0.841 × EVA + 0.134 × RUN − 0.069) with an R² of 0.868.

The five comparative experiments conducted here served to elucidate the intricate relationships between vegetation growth and various hydrological factors, shedding light on the inherent connections among these factors. Firstly, the basic linear regressions of the NDVI on ten hydrological factors provided a foundational understanding of how vegetation growth correlated with each individual factor and of which hydrological factors played significant roles in influencing vegetation growth. Secondly, the multiple linear regressions for the NDVI and two sets of hydrological factors further deepened understanding by exploring the comprehensive impacts of these factors on vegetation growth and offered insights into how different hydrological variables jointly impact vegetation dynamics. Moreover, the explorations of relationships between hydrological factors, such as the TWSA and TWSC, provided vital perspectives into the dynamic changes in water storage and distribution within ecosystems. Understanding these relationships was crucial for predicting and managing water resources in terrestrial environments.

4. Discussion

4.1. TWSA Trends Across Continents and Uncertainty of SSA

The observed global decline in the TWSA is consistent with ongoing hydrological changes influenced by both natural variability and anthropogenic activities (Table 1). The dramatic TWSA decrease in Greenland (−0.7262 cm/month) reflected accelerated glacier and ice-sheet melting due to rising temperatures, aligning with findings from Velicogna [81]. Similarly, the substantial decline in the TWSA in North America (−0.0764 cm/month) was attributed to glacier retreat in the Gulf of Alaska and the Canadian Archipelago [82,83]. Among the continents analyzed, the most remarkable facet of the changing TWSA was that terrestrial non-frozen water seemed to be accumulating in these tropical regions, including Africa and Oceania. In contrast, the significant losses were concentrated in the mid-latitudes, including Asia, North America, Europe, and South America, which is consistent with the work of Humphrey and Rodell [84,85]. In this century, the widespread decrease in precipitation in mid-latitudes, coupled with the general increase in high-latitude cold zones and low-latitude tropical regions, has led to a significant decrease in the TWSA in Eurasia and South America and a notable increase in Africa and Oceania [86,87,88]. However, the TWSA time series in South America showed obvious periodicity, and the series in Oceania basically remained stable but irregular, resulting in comparatively smaller TWS trends and lower coefficients of determination.

The TWSA time series across continents from 2011 to 2019 after SSA gap-filling are represented in Figure 12, demonstrating the effectiveness of the two stages in reconstructing missing data within the GRACE and GRACE-FO mission periods. SSA-filling-a was used to fill short-term gaps occurring within either the GRACE or GRACE-FO missions. In this stage, the optimal parameters M and L were uniformly determined prior to application across all regions. The reconstruction error for SSA-filling-a was quantified as the mean absolute residual between the observed time series and its SSA reconstruction across all valid time points, providing a measure of reconstruction uncertainty for existing data. In contrast, SSA-filling-b was employed to reconstruct the 11-month observational gap between the two missions. For this stage, parameters were fine-tuned through cross-validation: a full year within the observed period (2004–2016) was artificially masked, reconstructed using SSA, and compared with the original values. The reconstruction error was calculated as the average deviation across multiple such iterations and aggregated across all pixels in each region. Consequently, errors associated with SSA-filling-b appeared numerically larger, particularly over hydrological dynamic or data-sparse regions such as Greenland and Oceania. Overall, SSA-filling-a consistently exhibited smaller errors compared to SSA-filling-b, indicating higher reliability for stable portions of the time series. The uncertainties observed in SSA-filling-b primarily reflected the inherent challenges of reconstructing data across the extended mission gap under complex hydrological variability.

In this study, the SSA method was directly based on the mass concentration product from CSR for interpolation. Compared to the work of Yi and Sneeuw [21], who derived results in the form of spherical harmonic coefficients (SHCs), the mascon solutions not only enhanced signal resolution but also mitigated potential error sources such as unfiltered data errors and distorted RCs [89]. Therefore, the improved SSA method employed here was more convenient, efficient, and accurate, with errors falling within acceptable ranges. Overall, the SSA approach has been proven to be suitable for filling missing data from fluctuating patterns identified within existing long-term observations at a global scale [21,46].

4.2. Ecological Explanations of Differences in Different Vegetation Types

To reduce the uncertainty introduced by heterogeneous land cover and to ensure a clearer interpretation of NDVI-TWSA relationships, we concentrated our analysis on regions dominated by uniform vegetation types, thereby allowing more accurate ecological explanations of the observed correlations and lag patterns. As revealed by Figure 5, Figure 6, Figure 7, Figure 8 and Figure 9, distinct vegetation types exhibited varying spatial responses and lag patterns to the change in terrestrial water storage, primarily driven by differences in their distribution characteristics, hydrological contexts, phenological cycles, and rooting structures.

Tree cover occupied approximately 32.27% of the global terrestrial surface, encompassing both evergreen and deciduous forest types. These forests were predominantly distributed across temperate and subtropical regions such as Central South America, Russia, and southern Asia. Among these regions, around 24.47% exhibited a significant positive correlation between the NDVI and TWSA, particularly in southern Africa and southern Asia. These areas were characterized by warm climates, high levels of solar radiation, and pronounced seasonal variations in precipitation. As a result, vegetation growth was highly dependent on periodic hydrological replenishment, and increases in the TWSA directly enhanced vegetation productivity [90,91]. In contrast, more than 53% of global forested areas displayed a significant negative correlation between the NDVI and TWSA, especially in high-latitude areas such as northern North America and northern Eurasia. In these regions, vegetation growth was constrained primarily by temperature and solar radiation rather than by water availability. During the growing season, increased photosynthetic activity led to more water consumption, thereby depleting soil and subsurface moisture and resulting in a negative correlation between vegetation greenness and water storage [92]. Furthermore, forested areas typically exhibited a delayed NDVI response to TWSA changes, with observed time lags ranging from 8 to 11 months. Forests in North and South America predominantly responded after 8 to 9 months, whereas those in Eurasia and Africa displayed longer response times of 10 to 11 months. These extended lags reflected not only regional hydroclimatic regimes but also the intrinsic ecological characteristics of forest ecosystems. The deep and extensive root systems typical of forest vegetation allowed access to moisture in deeper soil layers, thereby reducing immediate dependence on surface water fluctuations and prolonging the response to hydrological changes [93,94]. Such a prolonged lag has also been highlighted by Liu [95], who reported asynchronous correlations between TWS declines and vegetation growth in forested areas, with time lags extending up to 13 months. A particularly notable case was observed in Russia, where vast forested regions displayed a spatial juxtaposition of very short (1–2 months) and very long (10–12 months) response lags. This abrupt spatial variability was likely a manifestation of strong climatic seasonality and annual vegetative cycles, which can obscure or distort the detection of continuous lag signals. Moreover, our approach may have limitations in accurately capturing lag durations that exceed one full phenological cycle, a constraint acknowledged in this study.

Shrubland ecosystems, which accounted for 6.36% of the global terrestrial surface, were primarily distributed across semi-arid regions such as eastern South America and southern Africa [90,96]. Around 85% of shrubland areas exhibited a significant positive correlation between the NDVI and TWSA, particularly in equatorial and subtropical zones characterized by high temperatures and evaporation rates. In these environments, vegetation growth was closely tied to short-term fluctuations in water availability, and increases in the TWSA resulting from episodic rainfall events effectively promoted shrub vegetation dynamics. The time lags between the TWSA and NDVI in shrub-dominated regions were notably short, typically within 1 to 2 months. This rapid response was attributable to the biological traits of shrubs, which generally possessed shallow but expansive root systems capable of swiftly accessing moisture in surface soils [97]. Moreover, the shorter life cycles and high phenological plasticity of shrub species further contributed to the prompt reaction to favorable hydrological conditions. These observations were consistent with the findings of Xie and Wu [24,27], who also reported immediate NDVI responses in semi-arid shrublands following precipitation events or increases in terrestrial water storage.

Grassland covered 23.4% of the global land area and was predominantly located within temperate continental and tropical savanna climates, including regions such as Australia, Kazakhstan, and western North America [98]. In these ecosystems, 45% of grassland areas exhibited a significant positive correlation between the NDVI and TWSA. Notable examples included extensive grasslands in Australia and southern Africa, where correlation coefficients exceeded 0.6, indicating that water availability directly influenced biomass accumulation. Conversely, approximately 30% of global grasslands, particularly in northern North America and Europe, showed a negative correlation. In these higher-latitude regions, vegetation growth was often constrained by low temperatures and limited solar radiation during shorter growing seasons. Under such conditions, increased vegetation activity frequently led to the rapid depletion of surface and subsurface moisture, resulting in a negative association with the TWSA. Grasslands exhibited intermediate time lags in vegetation response, generally ranging from 1 to 4 months. In southern Africa and Australia, lag durations were typically shorter (1 to 2 months), reflecting favorable hydrometeorological conditions and the efficient water uptake capacity of shallow-rooted grasses. By contrast, grasslands in Kazakhstan displayed longer lags (3 to 5 months), likely due to limited soil fertility and greater reliance on slowly recharging water sources. The shallow and dense root systems characteristic of grasses enabled rapid utilization of near-surface moisture, promoting swift but often short-lived growth responses to transient increases in water availability [91].

Cropland, which accounted for 8.77% of the global land, was primarily concentrated in regions under intensive anthropogenic management, such as India and the North China Plain [99]. Roughly 50% of cropland areas exhibited a significant positive correlation. For example, Indian croplands showed strong positive correlations, reflecting the high dependence of agricultural productivity on water availability, especially in areas experiencing groundwater depletion and reliant on monsoonal precipitation. Conversely, around 22% of croplands, predominantly in North America and Europe, demonstrated negative correlations. In these regions, water availability was generally sufficient, and crop growth was more constrained by other limiting factors such as solar radiation and temperature. Additionally, accelerated urbanization, coupled with advanced irrigation infrastructure and water management strategies, buffered the impacts of TWSA variability on vegetation dynamics [100]. Croplands responded relatively quickly to fluctuations in the TWSA, with time lags typically ranging from 1 to 4 months. This prompt response was attributed to their shallow root systems, well-drained soils, and short phenological cycles, which allowed for rapid exploitation of surface moisture. Furthermore, intensive agricultural practices, including irrigation and fertilization, further accelerated the vegetation response to changes in water storage [27]. In arid and semi-arid regions, such rapid adaptability was essential for maintaining crop resilience and ensuring food security.

Herbaceous wetlands, mangroves, and moss and lichen occupied only a small fraction of the global land and were spatially fragmented, warranting focused attention on their lag patterns and underlying causes. Herbaceous wetlands were predominantly located in moist environments such as marshes, wetlands, and riverbanks in western Russia. These ecosystems exhibited a significant negative correlation. Although characterized by shallow root systems, herbaceous wetland vegetation was highly sensitive to excessive moisture, with surplus water causing root asphyxiation and inhibiting growth [101]. Due to the intricate hydrological dependencies and extended growth cycles, NDVI responses typically lagged behind TWSA changes by around 8 months. Mangroves, covering only 0.10% of the global land area, were restricted to tropical and subtropical coastal regions [102]. Adapted to tidal environments through specialized root systems, mangroves exhibited a significant positive NDVI-TWSA correlation but remained vulnerable to drought and low temperatures [103]. The structural complexity of the ecosystems contributed to variable NDVI responses, generally with lags of 7–9 months. Moss and lichen, mainly distributed in cold and humid regions above 60°N such as the Canadian Archipelago and northern Russia [104], showed a significant negative correlation. Although they lacked root systems and relied on absorbing moisture from the atmosphere and surface layers, excessive moisture can lead to over-saturation and oxygen deficiency, which may inhibit photosynthesis and respiration. Their slow growth rates led to an approximate 8-month delay in the NDVI response. However, given their minimal reliance on subsurface water, future studies should incorporate surface soil moisture indices to better characterize the hydrological dynamics of these ecosystems.

In regions such as southern Africa, western Europe, and southern Asia, we identified sharp contrasts between short (1–2 months) and long (10–12 months) lags occurring in close spatial proximity. This interlacing pattern was due to the coexistence of multiple vegetation types. For instance, southern Africa included a mix of forest, shrubland, and grassland, while southern Asia and western Europe comprised cropland, forest, and grassland. Forests, with deep root systems and prolonged phenological cycles, exhibited delayed responses to changes in water availability. In contrast, croplands, grasslands, and shrublands, characterized by shallower root structures and shorter growth cycles, responded much more rapidly. These inter-vegetation contrasts gave rise to pronounced boundaries in the spatial distribution of lag durations. Moreover, our findings revealed that the NDVI responses to the TWSA exhibited considerable time lags in certain regions and vegetation types, which may span across years (≥12 months) or growing periods (≥7 months). Although these lags may seem inconsistent with typical hydrological responses, it was important to realize that the TWSA represented an integrated measure encompassing surface water, groundwater, and snow/ice water. While surface water typically has a rapid and immediate impact on vegetation growth, deeper water sources, including groundwater and soil moisture in deeper layers, should take much longer to recharge and become accessible to plants. These delayed contributions are particularly evident in arid and semi-arid regions, where water cycles and vegetation phenological processes are more complex. For instance, studies have shown that plants in dry regions often rely on groundwater or deep soil water for sustenance, especially during periods of limited rainfall. These deeper water sources are replenished over extended periods, leading to delayed impacts on vegetation growth [105,106]. Perennial plants and tree species with extensive root systems are particularly adept at accessing these deeper aquifers, allowing them to survive and respond to hydrological changes over months or even years [107]. Additionally, regions reliant on snowmelt or glacial contributions, such as parts of Asia and North America, experience a time lag between TWSA changes and water availability, as the melting and infiltration processes are inherently gradual [108]. These observations align well with the spatial patterns of lags identified in our study, providing a plausible explanation for the longer response times observed in certain regions.

In summary, this study revealed distinct spatial distributions and temporal response patterns among different vegetation types to terrestrial water storage anomalies. Tree cover exhibited time lags of approximately 9 months, reflecting sustained water uptake from deep soil layers and highlighting the importance of long-term conservation strategies for forests under drought conditions. Shrublands and grasslands responded rapidly, suggesting the potential for facilitating swift ecological restoration in degraded semi-arid regions. Croplands also demonstrated prompt responses to changes in water availability, underscoring the need for water-efficient agricultural planning. Ecosystems with longer response times, such as wetlands and mangroves, required hydrological management to maintain stability. Moss and lichen, due to their high sensitivity to moisture fluctuations, can serve as indicators of ecosystem shifts in vulnerable northern regions. Overall, understanding the various responses of vegetation to changes in terrestrial water storage can enhance predictive ecosystem modeling and support the development of sustainable land management strategies in the context of escalating climate challenges.

4.3. Limitations and Prospects

This study had certain limitations that constrained the depth of the analysis. The inclusion of deep groundwater and managed reservoirs in the TWSA data introduced noise, diminishing its relevance for accurately capturing vegetation dynamics. Moreover, the use of globally averaged time series resulted in temporally smoothed lag estimations, which tended to average out regional variability and obscure localized vegetation–water interactions. This approach may have masked small-scale variability and introduced uncertainties in attributing specific ecosystem responses, particularly when lag estimates were close to rounding thresholds. Additionally, the application of a multiple linear regression framework provided only a first-order assessment of the relationships between vegetation growth and hydrological variables, limiting the capacity to capture the complex and non-linear interactions between above- and below-ground processes.

Future research will focus on refining TWSA datasets by excluding deep groundwater components and managed reservoirs and prioritizing plant-accessible soil moisture layers to improve ecological relevance. Furthermore, dynamic land surface models and biophysical simulation frameworks will be incorporated to better represent vegetation growth processes and their responses to hydrological variability, explicitly accounting for interactions across different soil depths and vegetation structures. These methodological improvements will help validate and refine our understanding of the intricate interactions between hydrological changes and vegetation dynamics.

5. Conclusions

In this study, we analyzed the global spatiotemporal responses of vegetation growth to terrestrial water storage (TWS) using 20 years of GRACE/GRACE-FO and NDVI data. Our results indicated a pronounced global decline in TWS, with nearly half of the global land area affected, highlighting critical implications for water resource management under climate change. The relationship between vegetation growth and TWS varied significantly across regions and vegetation types. Positive responses were concentrated in low-latitude tropical areas, whereas negative correlations were observed in higher-latitude regions and tropical rainforests. Time-lag effects were evident, with shallow-rooted vegetation like grassland and shrubland responding rapidly, while deep-rooted vegetation such as tree cover showed delayed responses. Moreover, the bidirectional causality demonstrated that vegetation not only responded to hydrological changes but also influenced water cycles. These findings enhanced our understanding of the intricate water–vegetation dynamics and offered valuable guidance for ecological restoration and sustainable water management.