Exploration of Spatiotemporal Covariation in Vegetation–Groundwater Relationships: A Case Study in an Endorheic Inland River Basin

Abstract

Groundwater plays a vital role in sustaining dryland ecosystems, yet our understanding of the spatiotemporal dynamics of groundwater–vegetation interactions in endorheic river basins remains limited. In this study, the covariation between the normalized difference vegetation index (NDVI) and water table depth (WTD) in the Heihe River Basin (HRB), a representative endorheic system, is investigated via multisource data and generalized additive models (GAMs). The results indicate that the NDVI peaks in summer (July), with a corresponding decline in the WTD, indicating a basin-wide negative correlation. Spatial analysis reveals distinct upstream–downstream gradients: upstream regions exhibit strong seasonal synchronization, whereas midstream and downstream areas show weaker correlations because of mixed surface and groundwater influences. Landcover and climate significantly affect these interactions, with arid zones showing the strongest negative correlations (ρ = −0.38), particularly in wetlands, whereas humid regions show nonsignificant relationships. Geomorphological analysis highlights stronger correlations in mountainous areas than in low-relief plains. Positive correlations are the most prevalent in arid regions (54.5%), followed by hyper-arid regions (28.9%), while negative correlations also dominate arid regions (54.6%), followed by semiarid regions (27.6%). Cross-correlation analysis reveals synchronous NDVI–WTD changes at 95% of the grid points, with 5% exhibiting time lags (1–3 months), indicating localized hydrogeological feedback. Notably, 32% of the zones with negative correlations overlap with groundwater-dependent ecosystems (GDEs). GAM analysis reveals that 87.9% of the spatial variability in the NDVI–WTD correlations is attributed to environmental factors, with climate (26.6%) and hydrogeology (19.5%) as the dominant contributors. These findings provide critical insights into groundwater–vegetation interactions in arid ecosystems and offer valuable implications for sustainable water resource management.

1. Introduction

Water availability is critical for vegetation growth, with groundwater playing a fundamental role in sustaining ecosystems, agriculture, and human populations, particularly in arid and semiarid regions where surface water is limited [1,2,3,4]. Groundwater is essential for plant growth, and the response of vegetation to water availability influences the exchanges of water, carbon, and energy between the land and the atmosphere [5,6]. Vegetation dynamics are intricately linked to groundwater, with plant growth and distribution often reflecting the depth and availability of the water table [7,8,9]. In many ecosystems, particularly those in water-limited regions, vegetation depends on groundwater to meet water requirements, especially during dry periods [10,11]. This dependence highlights the importance of groundwater in maintaining ecological balance and supporting biodiversity. Therefore, the interaction between groundwater availability and vegetation dynamics is a critical area of study as it influences biodiversity, agricultural productivity, and the resilience of ecosystems to climate variability [12,13,14]. Understanding these interactions is essential for effective water resource management and conservation efforts [15,16].

Despite the growing recognition of groundwater–vegetation interdependence, previous studies have focused predominantly on static or large-scale assessments, often overlooking fine-grained spatiotemporal dynamics at the regional scale—a gap particularly pronounced in endorheic inland river basins, which exhibit unique vegetation–groundwater dynamics owing to terminal drainage and minimal hydrological connectivity [17,18,19]. Koirala et al. (2017) [7] mapped global groundwater–vegetation covariation using gross primary productivity data (10 km resolution) and groundwater model simulations, revealing significant correlations in 67% of vegetated areas but missing subbasin seasonal fluctuations critical in endorheic zones. Maxwell and Condon (2016) [5] quantified continental-scale transpiration partitioning via integrated hydrologic modeling, which improved accuracy by 15% with respect to lateral groundwater flow, but neglected spatial heterogeneity in earth system models. These limitations, particularly the coarse spatiotemporal resolution of global models that obscure seasonal feedback and the lack of consideration for spatial heterogeneity in groundwater-dependent vegetation responses, hinder the characterization of transient dynamics in groundwater–vegetation interactions, such as the relationship between the normalized difference vegetation index (NDVI) and water table depth (WTD) within endorheic basins [20,21,22,23,24,25,26]. Although hydrological simulations address local-scale interactions and remote sensing studies explore regional patterns [27,28,29], basin-scale understanding of the spatial extent, timing, and governing conditions of these interactions remains fragmented. This fragmentation limits the predictive capacity for ecosystem responses to climate extremes, such as prolonged droughts in arid regions [8,10,30,31].

In these closed systems, the absence of outflow pathways amplifies feedback loops between groundwater depletion and vegetation stress, yet global classifications often conflate endorheic basins with other groundwater-dependent ecosystems, masking their distinct vulnerability [32]. Vegetation here may temporarily stabilize water tables via hydraulic redistribution during droughts, but risks irreversible regime shifts if critical WTD thresholds are breached [33,34,35]. Current hydrological models oversimplify basin dynamics by assuming steady-state recharge rates, despite evidence that episodic rainfall and anthropogenic extraction drive nonstationary interactions [36,37]. Policy frameworks further compound the issue by ignoring basin-specific temporal lags between groundwater decline and vegetation collapse, while sparse long-term monitoring networks hinder the empirical validation of theoretical relationships [15,38]. Consequently, resolving spatiotemporal covariation in vegetation–groundwater relationships is essential to unravel feedback mechanisms and design adaptive management strategies for closed systems.

Moreover, understanding the spatial and temporal origins of plant water sources—including the relative contributions of shallow versus deep groundwater and seasonal shifts in precipitation uptake—is critical for refining Earth system models and optimizing ecosystem management [12,13,39,40]. Recent studies have highlighted that neglecting spatiotemporal variability in water sourcing undermines predictions of vegetation resilience to drought and groundwater depletion [26,41,42,43]. This highlights the need for integrated approaches that trace hydrological pathways, a gap this study addresses by combining high-resolution NDVI–WTD analyses with environmental covariates to disentangle groundwater dependency patterns [43,44].

The importance of groundwater in supporting vegetation growth is underscored by groundwater-dependent ecosystems (GDEs), which rely on subsurface water to sustain ecological functions in water-scarce regions [33,45,46,47]. GDEs are highly sensitive to changes in groundwater availability, whether from natural variability or anthropogenic extraction [11]. Identifying NDVI–WTD correlations and their spatial alignment with GDEs is crucial for sustainable management [14]. However, quantifying these interactions remains challenging in arid regions because of sparse vegetation and limited shallow groundwater [48,49,50]. A comprehensive understanding of NDVI–WTD spatiotemporal covariation is thus critical to mitigate the impacts of groundwater depletion and ensure resource sustainability in drylands.

A critical challenge in vegetation–groundwater dynamics lies in disentangling bidirectional interactions between vegetation and groundwater, particularly in endorheic basins, where environmental complexity obscures causation [51,52,53]. While Glanville et al. (2023) [9], in their systematic review of 140 global studies, identified methodological biases in correlative research (e.g., 50.7% of papers focused on arid regions, such as deserts and xeric shrublands, with 37.9% of studies analyzing desert ecosystems), their analysis highlighted unresolved causal mechanisms—such as conflating NDVI trends driven by groundwater uptake versus root adaptations—due to insufficient separation of these drivers in existing frameworks [39]. In addition, conventional remote sensing methods struggle to resolve subsurface processes at the plant–root scale in regions with dynamic water tables [54], whereas climatic variables (e.g., precipitation-driven soil moisture and temperature-modulated transpiration) further complicate vegetation–groundwater relationships. To address these gaps, we employ generalized additive models (GAMs), which excel in analyzing nonlinear spatiotemporal interactions within high-dimensional environmental systems [55,56,57]. By integrating penalized smoothing techniques to avoid overfitting [58], GAMs provide a robust framework to unravel NDVI–WTD covariation patterns, ultimately advancing ecological thresholds for sustainable groundwater management in water-scarce regions.

This study aims to (1) characterize the spatiotemporal coupling mechanisms between vegetation dynamics (NDVI) and groundwater fluctuations (WTD) in the Heihe River Basin (HRB), an endorheic inland basin, by quantifying seasonal synchrony, regional heterogeneity, and time-lagged interactions; (2) identify dominant environmental drivers (climate, soil, hydrogeology, vegetation traits, and human activities) governing the NDVI–WTD relationships through nonlinear modeling using GAMs; and (3) evaluate the spatial alignment between patterns of vegetation–groundwater covariation and mapped GDEs to inform adaptive water management strategies in arid regions.

2. Materials and Methods

2.1. Study Area

The HRB, a 143,000 km² endorheic basin in arid Northwest China, is an ideal natural laboratory for investigating spatiotemporal vegetation–groundwater interactions (Figure 1a,b) [59,60]. Spanning an elevation gradient from the Qilian Mountains (5500 m) to the Gobi Desert (<1000 m), the HRB encompasses three distinct subregions (Figure 1c): the upper reaches (uHRB, water source area with alpine meadows and forests), the middle reaches (mHRB, agricultural zone with intensive irrigation), and the lower reaches (lHRB, ecologically sensitive terminal area with desert vegetation and oases) [61,62,63]. This topo-climatic gradient drives sharp transitions in hydro-ecology, from humid alpine conditions (uHRB: 400–600 mm/y precipitation) to hyper-arid deserts (lHRB: <100 mm/y), creating a natural laboratory to study water-limited vegetation–groundwater linkages (Figure 1e) [64,65,66,67].

The HRB’s pronounced spatial heterogeneity in climate, landcover (nine types, dominated by deserts and groundwater-dependent oases; Figure 1d,f), and hydrogeology underpins its suitability for this study. Shallow groundwater (<5 m) in alluvial plains sustains riparian and agricultural ecosystems, whereas deep aquifers (>50 m) in desert margins impose critical constraints of water on vegetation [21,63]. Human activities—particularly irrigation in the mHRB and groundwater extraction in the lHRB—intensify vegetation–groundwater coupling, whereas natural gradients in soil permeability (thin alpine soils vs. alluvial deposits) regulate recharge patterns [68,69,70,71]. Supported by the HiWATER monitoring network [72], the integration of natural gradients and anthropogenic pressures in the HRB provides a unique framework to unravel how variations in the water table depth drive vegetation dynamics across scales—a central focus of this study [73].

2.2. Dataset

In this study, multiple datasets covering a range of environmental, climatic, hydrological, and vegetation-related variables are utilized to investigate groundwater–vegetation interactions. These datasets are obtained from publicly available sources and undergo preprocessing, including reprojection, resampling, and clipping, to ensure consistency across spatial and temporal scales. All datasets are projected onto a uniform coordinate system and resampled to a consistent spatial resolution of 0.005°, ensuring compatibility and comparability across different environmental variables. A brief summary is shown in Table 1, and detailed information is provided in Table S1.

NDVI data at a 30 arc-second resolution are derived from cloud-free MODIS data spanning 2001–2011. These products, which are based on MODIS (MYD13A2 and MOD13A2) data, use the Harmonic Analysis of Time Series (HANTS) algorithm to remove clouds and reconstruct the daily resolution NDVI dataset [74]. WTD data are obtained from a global groundwater dataset at a 30 arc-second resolution and represent groundwater conditions from 2003 to 2013 [41]. The climate classification dataset is obtained from a global climate zoning dataset, the Global Aridity Index and Potential Evapotranspiration Database—Version 3 (Global-AI_PET_v3) [75], which categorizes regions into different climate types on the basis of long-term meteorological variables. Landcover types are obtained from the 30 m 5-yearly landcover maps of the Qilian Mountain area, which include nine categories with an overall accuracy exceeding 90% [76]. The dataset of geomorphological types is obtained from the Resource and Environmental Science Data Platform (https://www.resdc.cn/, accessed on 10 January 2025), providing detailed information on regional landform characteristics, including those of plains, hills, and mountains. The classification criteria for geomorphology are provided in [7].

Table 1. Summary of the datasets used in this study. For additional details, please refer to Supplementary Information (Table S1).

Parameter/Variable	Acronym	Original Spatial		Original Temporal		Source
		Domain	Resolution	Period	Resolution
Climate class	-	Global	30 arc-second	1960–2000	-	[75]
Landcover type	-	Qilian Mountain Area	30 m	1990–2020	5-year	[76]
Geomorphology type	-	China	1 km	-	-	-
Water table depth	WTD	Global	30 arc-second	2003–2013	Monthly	[41]
Normalized difference vegetation index	NDVI	HRB	0.01 degree	2001–2011	Daily	[74]
Precipitation	P	China	0.025 degree	1951–2011	Monthly	[77]
Land surface temperature	LST	China	1 km	2000–2022	Daily	[78]
Photosynthetic effective radiation	PAR	HRB	1 km	2010–2015	Hourly	[79]
Consecutive dry days	CDD	Global	0.1 degree	-	-	[80]
Plant-available soil water	PASW	Global	0.5 degree	-	-	[81]
Soil organic matter	SOM	China	30 arc-second	-	-	[82]
Soil pondus hydrogenii	pH	China	30 arc-second	-	-	[82]
Soil texture	ST	Global	30 arc-second	-	-	[83]
Maximum rooting depth	MRD	Global	30 arc-second	2003–2013	Monthly	[84]
Leaf area index	LA	HRB	30 m	2011–2015	Monthly	[85]
Gross primary productivity	GPP	Global	1 km	2001–2023	Monthly	-
Growing season length	GSL	Global	300 m	-	-	[86]
Porosity	Ps	Global	Vector	-	-	[87]
Saturated hydraulic conductivity	Ks	Global	Vector	-	-	[87]
Aquifer thickness	AT	Global	3 arc-second	-	-	[88]
Vadose zone thickness	VZT	Global	30 arc-second	-	-	[89]
Digital elevation model	DEM	Qilian Mountain area	30 m	2018	-	[90]
Topographic index	TI	Global	30 arc-second	-	-	[91]
Human modification	HM	Global	30 arc-second	2016 1	-	[92]
Groundwater-dependent ecosystem	GDE	Global	30 m	-	-	[93]

¹ Here, the temporal range has a median year of 2016.

Table 1. Summary of the datasets used in this study. For additional details, please refer to Supplementary Information (Table S1).

Parameter/Variable	Acronym	Original Spatial		Original Temporal		Source
		Domain	Resolution	Period	Resolution
Climate class	-	Global	30 arc-second	1960–2000	-	[75]
Landcover type	-	Qilian Mountain Area	30 m	1990–2020	5-year	[76]
Geomorphology type	-	China	1 km	-	-	-
Water table depth	WTD	Global	30 arc-second	2003–2013	Monthly	[41]
Normalized difference vegetation index	NDVI	HRB	0.01 degree	2001–2011	Daily	[74]
Precipitation	P	China	0.025 degree	1951–2011	Monthly	[77]
Land surface temperature	LST	China	1 km	2000–2022	Daily	[78]
Photosynthetic effective radiation	PAR	HRB	1 km	2010–2015	Hourly	[79]
Consecutive dry days	CDD	Global	0.1 degree	-	-	[80]
Plant-available soil water	PASW	Global	0.5 degree	-	-	[81]
Soil organic matter	SOM	China	30 arc-second	-	-	[82]
Soil pondus hydrogenii	pH	China	30 arc-second	-	-	[82]
Soil texture	ST	Global	30 arc-second	-	-	[83]
Maximum rooting depth	MRD	Global	30 arc-second	2003–2013	Monthly	[84]
Leaf area index	LA	HRB	30 m	2011–2015	Monthly	[85]
Gross primary productivity	GPP	Global	1 km	2001–2023	Monthly	-
Growing season length	GSL	Global	300 m	-	-	[86]
Porosity	Ps	Global	Vector	-	-	[87]
Saturated hydraulic conductivity	Ks	Global	Vector	-	-	[87]
Aquifer thickness	AT	Global	3 arc-second	-	-	[88]
Vadose zone thickness	VZT	Global	30 arc-second	-	-	[89]
Digital elevation model	DEM	Qilian Mountain area	30 m	2018	-	[90]
Topographic index	TI	Global	30 arc-second	-	-	[91]
Human modification	HM	Global	30 arc-second	2016 1	-	[92]
Groundwater-dependent ecosystem	GDE	Global	30 m	-	-	[93]

¹ Here, the temporal range has a median year of 2016.

Additionally, a range of environmental variables related to climate, soil, vegetation, hydrogeology, topography, and human activity are incorporated both to quantify the contributions of these factors to the NDVI–WTD relationship and to assess their influence on vegetation–groundwater interactions. The climatic factors considered in this study include precipitation (P), land surface temperature (LST), photosynthetically active radiation (PAR), and consecutive dry days (CDD), which collectively regulate the water availability, energy balance, and plant physiological responses. The soil characteristics include potential plant-available soil water (PASW), soil organic matter (SOM), soil pondus hydrogenii (pH), and soil texture (ST), each of which plays crucial roles in governing soil–water interactions, nutrient cycling, and plant growth conditions. The vegetation attributes include the maximum rooting depth (MRD), leaf area index (LAI), gross primary productivity (GPP), and phenological stage, represented by the growing season length (GSL), all of which are essential indicators of ecosystem productivity and vegetation–climate feedback. The hydrogeological parameters evaluated in this study include saturated hydraulic conductivity (Ks), porosity (Ps), vadose zone thickness (VZT), and aquifer thickness (AT), which determine groundwater movement, recharge capacity, and subsurface storage dynamics. Additionally, topographic features, represented by a digital elevation model (DEM) and topographic index (TI), influence surface runoff, groundwater flow paths, and landscape hydrological connectivity. Finally, human activity is quantified through human modification (HM), which reflects anthropogenic alterations to natural hydrological and ecological systems.

P data are derived from a gridded dataset with a spatial resolution of 0.025°, providing monthly precipitation estimates from 1951 to 2011 [77]. LST data are obtained from a 1 km daily all-weather dataset covering the period from 2000 to 2022 [78]. Photosynthetically active radiation (PAR) data are sourced from an hourly resolution dataset spanning 2010 to 2015 [79]. CDD data are extracted from a global climate dataset with a spatial resolution of 0.1°, offering long-term drought condition assessments [80]. The PASW is extracted from global maps of potential and climatic plant-available soil water, with a spatial resolution of 0.5° [81]. SOM and soil pH data are sourced from the Soil Database of China for Land Surface Modeling [82]. ST is extracted from a high-resolution global dataset called “A new version of the global high-resolution dataset of soil hydraulic and thermal parameters for land surface modeling” [83], which characterizes soil composition across different regions. Here, the clay content is used to present the ST. MRD data, which are obtained from a global plant rooting depth [84], are used to estimate the depth of root penetration of vegetation. The LAI data are sourced from a 30 m monthly composite dataset, which represents vegetation structural characteristics [85]. The GPP data are available at https://www.resdc.cn/data.aspx?DATAID=204, accessed on 8 January 2025. The GSL is extracted from a 300 m global dataset, providing information on phenological variations [86]. Ks and Ps data are obtained from a high-resolution global hydrogeological dataset, the GLobal HYdrogeology MaPS (GLHYMPS) [87]. AT is represented by the depth to bedrock, for which geological survey data, borehole measurements, and lithostratigraphic modeling are synthesized to estimate the depth and extent of aquifer systems worldwide [88]. The VZT dataset used in this study is derived from a global-scale hydrogeological model that integrates soil properties, lithological classifications, and water table observations [89]. DEM data are sourced from a 30 m ASTER-GDEM dataset [90]. The topographic index dataset is generated from the HydroSHEDS database [91]. Human modification data, which represent anthropogenic impacts on terrestrial ecosystems, are sourced from the Global Human Modification dataset, which integrates 13 stressors to quantify human influence [92]. A global GDE map with a 30 m resolution is incorporated to identify regions where vegetation relies on groundwater resources [93]. The spatial distributions of these environmental variables are illustrated in Figure 2, providing a comprehensive characterization of the key drivers influencing groundwater–vegetation interactions and hydrological processes across the HRB.

2.3. Methods

In this study, multisource, high-resolution datasets (e.g., remote sensing data, modeled outputs, GIS-processed spatial layers, ground-based measurements) that combine model-based WTD estimates with satellite-derived NDVI data are utilized to investigate the spatial and temporal relationships between vegetation dynamics and groundwater fluctuations at a regional scale. The spatiotemporal variations in the NDVI and WTD are analyzed across different months, and the characteristics of distinct subregions are examined in detail. Additionally, reciprocal influences between vegetation and groundwater are assessed to determine how vegetation responds to groundwater availability and vice versa. To quantify the strength and patterns of vegetation–groundwater relationships, Spearman’s rank correlation and cross-correlation analyses are employed. These statistical methods allow for the evaluation of the spatial distribution of the NDVI–WTD interactions and the assessment of temporal alignment or lag effects between the NDVI and WTD variations. Furthermore, this study examines how these spatiotemporal patterns are influenced by climate classifications, landcover types, and geomorphological characteristics. In addition to evaluating broad environmental influences, this study investigates the relationships between the NDVI–WTD correlations and GDEs. The contributions of key environmental variables to the NDVI–WTD relationships are quantified using GAMs, which allows the identification of nonlinear effects across multiple environmental factors. The workflow and methodology of this study are depicted in Figure 3. A concise overview of the methodologies employed in this study is provided below.

2.3.1. Spearman’s Rank Correlation Analysis

Spearman’s rank correlation coefficient (ρ) is employed to assess the monotonic relationship between the NDVI and WTD across different grid cells. Unlike Pearson’s correlation, which measures linear associations, Spearman’s correlation evaluates the strength and direction of a monotonic relationship by ranking the values before computing the correlation. The coefficient is defined as follows:

$$\rho =1-{\displaystyle \frac{6\sum d_i^2}{n\left(n^2-1\right)}}$$

(1)

where d_i is the rank difference between paired NDVI and WTD values and n is the number of valid observations. Hydrological and vegetation processes frequently exhibit nonlinear relationships, making Spearman’s rank correlation particularly suitable for capturing these complex interactions. To investigate the spatiotemporal distribution of vegetation–groundwater relationships, we employ Spearman’s rank correlation and cross-correlation analyses. While Pearson and Kendall correlation coefficients are also calculated to assess the NDVI–WTD relationship (see Figure S4), Spearman’s coefficient shows the largest range of values, superior robustness, and interpretability. Consequently, Spearman’s correlation is adopted as the primary metric for subsequent analysis.

In the analysis, the NDVI and WTD time series are extracted for each grid cell, with no values removed to ensure valid computations. If fewer than six valid observations remain, the correlation is not calculated. Otherwise, Spearman’s correlation (ρ) and its corresponding p value are computed using MATLAB’s corr function. However, Spearman’s correlation coefficient shows the largest range of values and meaningful results. Given the complex and potentially nonparametric nature of the NDVI–WTD relationship, Spearman’s method is determined to be the most suitable for subsequent analysis.

2.3.2. Cross-Correlation Analysis

Cross-correlation is a statistical method used to assess the relationship between two time series, such as the NDVI and WTD, at different time lags. It measures how one series (e.g., NDVI) is correlated with a lagged version of another series (e.g., WTD), which is useful for identifying potential temporal dependencies and lead–lag relationships between two variables. The cross-correlation function (CCF, C_τ) can be defined as follows:

$$C_{\tau }={\displaystyle \frac{{\sum }_t\left(X\left(t\right)-{\mu }_X\right)\left(Y\left(t+\tau \right)-{\mu }_Y\right)}{\sqrt{{\sum }_t{\left(X\left(t\right)-{\mu }_X\right)}^2{\sum }_t{\left(Y\left(t+\tau \right)-{\mu }_Y\right)}^2}}}$$

(2)

where X(t) and Y(t) are the monthly time series of the NDVI and WTD, respectively, at time t, μ_X and μ_Y are their respective means, and τ represents the lag between the two series. The output of the cross-correlation function is a normalized correlation coefficient, which ranges from −1 (perfect negative correlation) to +1 (perfect positive correlation), with values closer to 0 indicating no correlation. The analysis is carried out after ensuring that both the NDVI and WTD datasets are cleaned of missing and null values and that the time series are adjusted to remove any invalid entries. The resulting cross-correlation values, which are displayed against the lag values, help in understanding the temporal relationship between vegetation dynamics and groundwater depth, revealing potential delays or lead effects in the response of vegetation to changes in the water table.

2.3.3. Generalized Additive Model (GAM)

The generalized additive model (GAM) is a flexible statistical framework that extends generalized linear models (GLMs) by incorporating smooth functions to model nonlinear relationships between predictor and response variables [94]. A GAM can be expressed as follows [95]:

$$g\left(E\left(Y\right)\right)={\beta }_0+f_1\left(X_1\right)+f_1\left(X_1\right)+\cdots +f_p\left(X_p\right)+ϵ$$

(3)

where g is the link function, E(Y) is Spearman’s rank correlation coefficient (ρ) between the NDVI and WTD, β₀ is the intercept, f_i(X_i) are smooth functions of the explanatory variables, and ϵ is the error term. These smooth functions allow for flexible, data-driven estimations of relationships, making GAMs particularly suitable for capturing complex nonlinear dependencies. GAMs utilize flexible smoothing functions—such as thin plate or cubic regression splines—to capture nonlinear trends without imposing a rigid, predetermined functional form, thus allowing the model to adapt to the intrinsic structure of the data. This adaptability is particularly crucial when investigating vegetation–groundwater relationships, where factors such as the NDVI and WTD often exhibit intricate and heterogeneous patterns spatially and temporally.

In this study, the GAM is applied using the “mgcv::gam” function in R to analyze the contributions of climate, soil, vegetation, hydrogeology, topography, and human activity to the NDVI–WTD relationship. The smoothing functions, often implemented as penalized splines, help mitigate overfitting while preserving interpretability. By leveraging the GAM, this study provides a robust, data-adaptive framework for understanding hydrological–vegetation interactions. For comprehensive details on the GAM, please refer to https://www.rdocumentation.org/packages/mgcv/versions/1.9-1/topics/gam, accessed on 1 May 2024. The application of the GAM in the Earth sciences, particularly in remote sensing and hydrology applications, has been extensively documented in the literature [55,96,97,98], demonstrating its utility in modeling complex nonlinear environmental processes.

3. Results

This section begins by presenting examinations of the seasonal dynamics of vegetation–groundwater relationships, followed by an assessment of their spatiotemporal distributions via Spearman’s rank correlation and cross-correlation analyses. Next, the variability in the NDVI–WTD correlations is evaluated across different landcover types and climate zones, along with their respective influences. Further analysis explores the role of geomorphological characteristics and their impact on vegetation–groundwater interactions. Additionally, we investigate associations with GDEs. Finally, the contributions of the environmental variables to the NDVI–WTD correlations are quantified using the GAM.

3.1. Seasonal Dynamics of Vegetation–Groundwater Relationships

We first explore the spatial and temporal variations in the NDVI and WTD in the HRB across different months using an NDVI–WTD covariation grid map. Figure 4 presents the temporal dynamics of the NDVI and WTD, illustrating the seasonal variability and spatial heterogeneity in groundwater–vegetation interactions. The trends in the WTD and NDVI exhibit distinct spatial patterns across the HRB, reflecting variations in hydroclimatic and geomorphological conditions. Basin-wide (Figure 4a), the NDVI shows a steady increase during the growing season (from March to August), with a peak in July, followed by a gradual decline in the cooler months. This pattern suggests a strong seasonal influence on vegetation growth. On the other hand, the WTD shows an inverse pattern, with a gradual increase in depth from the spring to winter months. The basin-wide trends reveal a negative correlation between the NDVI and WTD.

Specifically, in the upstream region (uHRB, Figure 4b), both the NDVI and WTD show pronounced seasonal variations, with the NDVI peaking in midsummer (June–July) as the WTD decreases, indicating high groundwater availability that supports vegetation growth. In contrast, the midstream area (mHRB, Figure 4c) shows less fluctuation in the WTD, which remains relatively deep throughout the year, suggesting that vegetation growth here is less dependent on groundwater and more influenced by a combination of groundwater and surface water, such as snowmelt or precipitation. The downstream region (lHRB, Figure 4d) displays minimal seasonal variation in both the NDVI and WTD, with consistently deeper water tables, indicating reduced groundwater availability and suggesting that vegetation growth is more reliant on surface water and climatic conditions than on groundwater. These regional differences reflect the influences of groundwater recharge, surface water availability, and climatic conditions on vegetation dynamics across the HRB.

To evaluate the reciprocal influences between vegetation dynamics and groundwater fluctuations, we analyzed the spatiotemporal variations using monthly average data. Figure 5 presents the monthly NDVI and WTD trends, annual mean NDVI, and seasonal and monthly spatial distributions of the NDVI (right panels). Figure 5c is the legend for Figure 5a,b. The colored matrix in Figure 5c displays the NDVI against the logarithm of the WTD, where the green regions indicate higher NDVI values associated with shallow WTDs, and the red regions represent areas with deeper water tables and lower NDVI values.

Figure 5a presents the annual mean variation in the NDVI and WTD across the study area, with a color gradient representing the varying depths. This map shows regions of shallower and deeper groundwater, providing a broad overview (baseline) of the spatial distribution of WTD across the basin. Shallower WTD is typically concentrated in certain areas, which likely correspond to regions with more accessible groundwater supporting vegetation growth, whereas deeper WTD is found in areas where groundwater is less accessible for plant uptake.

Figure 5b displays the monthly variations in the NDVI and WTD over the HRB, as well as the dynamics and NDVI–WTD interrelationships. During winter, most of the region has low NDVI values, corresponding to dormant vegetation. Both the NDVI and WTD show relatively stable patterns. This period is characterized by deeper water tables, indicating limited vegetation activity and groundwater uptake. The absence of significant green areas suggests minimal shallow WTD regions with active vegetation growth.

With the onset of spring, the NDVI begins to increase, particularly in southern and lower-elevation areas, as indicated by the expansion of yellow and green regions. Notably, the most significant NDVI improvements occur in the mHRB and uHRB, especially in the uHRB, where the NDVI and WTD exhibit a synchronous increase (positive correlation), suggesting coordinated vegetation growth and groundwater recharge. This finding suggests enhanced vegetation growth in areas where the WTD is relatively shallow, thereby promoting water availability for plant uptake. However, a significant portion of the landscape remains red, indicating persistent deep WTD (>60%) and lower NDVI (<40%) values. The correlation between the NDVI and WTD strengthens as vegetation growth becomes more responsive to water tables.

Summer exhibits the most pronounced vegetation activity, with extensive green and yellow areas, indicating widespread high NDVI values associated with both shallow and high WTD conditions. In the agricultural areas of the mHRB, widespread synchronous increases in the NDVI and WTD (positive correlations) reflect irrigation-driven groundwater use and crop growth. Moreover, lHRB regions with concurrent NDVI–WTD increases are confined primarily to desert oases, where rising water tables support vegetation resilience. This period represents peak vegetation productivity, with a strong negative NDVI–WTD correlation in regions where groundwater is a primary water source. However, some areas with deep WTD (red) remain, likely due to low water retention capacity or groundwater depletion.

As autumn progresses, the NDVI decreases, particularly in October and November, as vegetation senescence begins. The upstream areas show prominent synchronous declines in the NDVI-WTD (blue clusters), indicating coupled reductions in vegetation strength and the groundwater table. Moreover, emerging green patches in some regions signal decoupled NDVI–WTD dynamics. The green and yellow areas gradually shrink, whereas the red regions expand, indicating a return to lower NDVI values and deeper water tables. The correlation between the NDVI and WTD weakens in some areas as the dependency of vegetation on groundwater diminishes with increasing seasonal precipitation and cooling temperatures.

Figure 5d presents the statistical analysis derived from the data shown in Figure 5b, showing the monthly distributions of the NDVI–WTD combinations and highlighting the seasonal dynamics of these interactions. The analysis categorizes the data into four distinct groups: (1) high WTD (>60%) and low NDVI (<40%) values, represented in red; (2) low WTD (<40%) and high NDVI (>60%) values, shown in green; (3) low WTD (<40%) and low NDVI (<40%) values, depicted in blue; and (4) high WTD (>60%) and high NDVI (> 60%) values, indicated in yellow. The proportion of areas with high WTD and high NDVI (yellow) values increases notably during the growing season, reaching a peak in June (6%) and July (7%) before gradually declining. This trend aligns with the enhanced NDVI–WTD correlation observed in summer, suggesting that vegetation growth is closely linked to groundwater availability during peak productivity periods. The dominance of red regions throughout the year further emphasizes that most areas experience deep WTD and low NDVI values, reinforcing the spatial variability of vegetation–groundwater interactions.

Furthermore, Figures S2 and S3 present directional variograms (relevant instructions are recorded in Supplementary Text S1 and Figure S1) for the NDVI and WTD dynamics, revealing increased spatial autocorrelation with lag distance, which is particularly strong in summer (June–July) than in winter (December–January). The NDVI results in steeper seasonal variance gradients and pronounced summer spatial clustering, whereas the WTD results in gradual growth in variance with minimal directional seasonality, especially in diagonal orientations. The distinct monthly peak patterns between the NDVI and WTD highlight their divergent spatiotemporal structures, necessitating further mechanistic analysis.

3.2. Spatiotemporal Distribution of Vegetation–Groundwater Relationships

Figure 6a presents a map of the Spearman correlation between the NDVI and the WTD across the HRB. The color scale indicates the strength and direction of the correlation, ranging from −1 (strong negative correlation) to 1 (strong positive correlation). In the map, areas with positive correlations are marked in orange, and those with negative correlations are marked in blue. The variation across the basin highlights the complex interaction between groundwater dynamics and vegetation, with some areas showing strong positive and negative correlations and others showing no significant relationship. For example, sparsely vegetated or nonvegetated zones (e.g., deserts or impervious surfaces) show negligible NDVI–WTD correlations. Figure 6b offers a categorical analysis of the NDVI–WTD correlation, classifying regions into positive (ρ > 0.4, p value < 0.05), negative (ρ < −0.4, p value < 0.05), and insignificant (−0.4 < ρ < 0.4 or p value > 0.05) regions. The positive correlation (shown in orange) predominates in areas with low vegetation cover. Negative correlations (blue) are concentrated in areas with deeper water tables, such as mountainous regions, or areas with limited groundwater recharge.

Figure 6c–h shows magnified images of specific locations within the HRB, marked by labels d to h, to further analyze the spatial variability of the NDVI–WTD relationship. These locations are highlighted in Figure 6b to provide a closer look at areas with different correlation characteristics. Each marked location represents a specific geographic point, where distinct seasonal and spatial dynamics can be observed.

Specifically, Figure 6c–e present monthly time series data for the NDVI and WTD at locations ‘d’, ‘e’, and ‘f’ (as labeled in Figure 6b). Figure 6c corresponds to location ‘c’, showing a positive correlation (ρ = 0.92) between the NDVI and WTD. Here, the NDVI peaks during the growing season (June to August), aligning with a deep water table, indicating a strong dependency of vegetation on groundwater during this period. In Figure 6d, at location ‘d’, the correlation increases (ρ = 0.97), further reinforcing the relationship between shallow water tables and high vegetation productivity. This suggests that groundwater significantly supports vegetation growth during the peak season. However, in Figure 6e, at location ‘e’, the correlation remains positive (ρ = 0.92), indicating a more complex interaction where groundwater still supports vegetation, but other factors likely contribute to vegetation growth during the summer months.

Figure 6f–h depict the negative correlations between the NDVI and WTD, showing regions where deeper water tables coincide with lower vegetation activity. At location ‘f’ (Figure 6f), the negative correlation (ρ = −0.97) suggests that in areas with deeper groundwater, vegetation activity is suppressed, likely due to reduced groundwater availability. Similarly, Figure 6g (location ‘g’) also shows a high negative correlation (ρ = −0.91), reflecting a scenario where deeper water tables correlate with lower NDVI values. In Figure 6h, at location ‘h,’ the negative correlation (ρ = −0.96) continues to emphasize that areas with deeper groundwater depths tend to have less vegetation, possibly due to a lack of accessible water for transpiration. These negative correlations further highlight the critical role that shallow groundwater plays in supporting vegetation, particularly in arid regions, where groundwater is a primary source of water for ecosystems.

Figure 7a provides a pie chart illustrating the proportion of positive, negative, and insignificant correlations between the NDVI and WTD over the HRB. A total of 42.3% of the correlations are positive, 7.9% are negative, and 49.8% are not statistically significant. Figure 7b further details the frequency distributions of these correlation values. The histogram shows the frequency of the Spearman correlation coefficients for the positive, negative, and insignificant categories. The positive correlations are skewed toward higher values, with a mean of 0.572. The negative correlations are concentrated around lower values, with a mean of −0.595. The nonsignificant correlations, with a mean of 0.141, correspond to regions where groundwater depth has a negligible or no discernible effect on vegetation dynamics, implying that other factors may influence vegetation growth in these areas.

Then, a cross-correlation analysis is conducted to assess the temporal alignment or lag between the NDVI and WTD variations across the HRB. Figure 8a depicts the distribution of the time lags. The majority of the grids (95%) exhibit no time lag (time lag = 0), suggesting synchronous changes in vegetation and the water table. Among the remaining 5% (10,346 grids), positive lags (1–8 months) constitute 88% of the nonzero cases, suggesting that WTD responds to NDVI changes, whereas negative lags (−8 to −1 months) account for 12%, indicating the reverse pattern. Short-term lags (1–3 months) dominate, representing 58% (positive) and 9% (negative) of all lags, reflecting spatial heterogeneity in regional NDVI–WTD variations. The distribution of nonzero time lags reveals spatial heterogeneity in the temporal coupling between vegetation dynamics and groundwater fluctuations, which is likely driven by variations in soil properties, climatic conditions, and vegetation characteristics. Overall, the NDVI and WTD exhibit synchronized changes in most regions, with deviations occurring in a smaller subset, where the response of groundwater to vegetation shifts with a short delay.

To explore the hydro-pedological drivers of time-lagged vegetation responses, we conduct a comparative analysis of soil and hydrogeological parameters between the entire HRB and regions exhibiting a WTD lagging the NDVI (referred to as lagged-response regions). A summary table (Figure 8b) presents the mean values and ranges for eight key parameters: PASW, SOM, pH, ST, K, VZT, AT, and Ps. Spatial statistics are derived by masking valid pixels within the lagged-response regions on the basis of the time–lag matrix. The results revealed that the PASW (128.31 mm vs. 106.14 mm) and SOM (0.052 vs. 0.017) are significantly greater in the lagged-response region than in the entire HRB, whereas the pH (7.4 vs. 8.1), ST (clay content, 14.8 vs. 22.3), VZT (9.3 m vs. 17.9 m), and AT (16.6 m vs. 28.2 m) are notably lower. The higher PASW and SOM values in lagged-response regions likely improve soil water retention, extend the vertical movement of water, and contribute to delayed groundwater responses. In contrast, a reduced clay content may impede the efficiency of capillary rise, and thinner aquifers, along with limited water table fluctuations, may restrict rapid hydraulic adjustments to vegetation water demand. These factors collectively help explain the observed time-lagged coupled vegetation–groundwater patterns.

3.3. Variability in the NDVI–WTD Correlations Across Landcover Types and Climate Zones

Figure 9 presents boxplots showing the correlations between the NDVI and WTD across various landcover types (panel a), climate classes (panel b), and their interactions (panel c). Panel d provides a heatmap displaying the correlation values between landcover types and climate zones. Finally, panel e shows the distribution of the statistical correlation values across the different combinations of landcover and climate types, showing differences between positive, negative, and nonsignificant relationships.

Figure 9a shows the correlation between the NDVI and WTD variations across landcover types. Forests and grasslands show the strongest positive correlation, whereas croplands, shrublands, and wetlands exhibit negative correlations. Impervious surfaces and snow/ice areas show minimal variability, reflecting a more stable groundwater regime. Figure 9b shows that the relationship between the NDVI and WTD also varies across different climate classes. The hyper-arid zone shows the strongest positive correlation between the NDVI and WTD. In contrast, more humid climates (e.g., the humid zone) show a reduced or more neutral relationship, likely due to higher annual precipitation and more stable groundwater recharge processes that diminish the need for vegetation to rely heavily on shallow groundwater.

The combined effects of landcover and climate zones are analyzed in Figure 9c. The correlations between the NDVI and WTD in wetlands show significant variability across different climatic regions. Strong positive correlations are observed primarily in dry sub-humid areas with grasslands. In contrast, the most pronounced negative correlations occur in arid regions with wetlands, indicating a typical GDE synergy between vegetation and groundwater availability. In semi-arid and dry sub-humid regions, correlations are more heterogeneous, with some landcover types displaying weak relationships. However, positive correlations remain common in forests and grasslands across all climate classes, whereas croplands, wetlands, and shrublands present more variable or negative correlations.

To capture a comprehensive understanding of the relationship between the NDVI and WTD across various landcover types and climate classes, a correlation heatmap is utilized (Figure 9d). This shows substantial spatial variability in groundwater–vegetation dynamics. In semi-arid regions, the most robust positive correlations (approximately 0.33) are found in forests, with moderate correlations also observed in grasslands. In contrast, arid zones exhibit significant negative correlations, particularly in wetlands, where the correlation reaches approximately −0.38. Hyper-arid wetlands show weaker negative correlations (~−0.17), whereas bare lands exhibit positive correlations (approximately 0.29). In semi-arid wetlands, the correlation approaches zero, likely because of the influence of other hydrological factors. Landcover types, such as impervious surfaces and snow/ice, consistently show weak or neutral correlations across all climate zones, indicating limited interactions with groundwater.

Finally, the distribution of the statistical correlation values is depicted in Figure 9e, which shows that positive correlations are most prevalent in arid regions (54.5%), followed by hyper-arid regions (28.9%) and semi-arid regions (15.7%). Conversely, negative correlations are more prominent in semi-arid regions (27.6%) than in arid regions (54.6%) and hyper-arid regions (17.2%). This suggests that hyper-arid regions are characterized by positive correlations, whereas semi-arid regions are dominated by negative correlations.

In summary, the results suggest that in arid and semi-arid regions, groundwater availability plays a crucial role in sustaining vegetation. In hyper-arid zones, vegetation types, such as shrubland areas and wetland areas, are negatively correlated with water table depth, emphasizing the importance of groundwater in these water-scarce areas. However, in more humid climates, groundwater becomes less critical for vegetation growth, as surface water inputs from precipitation are likely the dominant factor influencing vegetation health.

3.4. Variability in the NDVI–WTD Correlations with Geomorphological Characteristics

Figure 10 presents the relationships between the NDVI and WTD across different geomorphological types (panel a) and water table classes (panel b). The classification criteria for both geomorphology and the water table are provided in Tables S4 and S5. The classification of geomorphological types is based on altitude relief, and six categories are distinguished [99]: plain (<30 m), platform (>30 m), hill (<200 m), low-relief mountain (200–500 m), middle-relief mountain (500–1000 m), and high-relief mountain (1000–2500 m). Plains (46.802%) and hills (25.189%) dominate, whereas middle- and high-relief mountains constitute smaller portions. The water table classification is based on depth ranges: shallow (0–2 m), relatively shallow (2–10 m), intermediate (10–30 m), relatively deep (30–100 m), and deep (>100 m). The largest category is intermediate-depth WTD (44.245%), followed by relatively deep WTD (23.940%), with the smallest category being deep WTD (4.608%).

The variations in the NDVI–WTD correlations across different landforms are shown in Figure 10a, indicating the influence of terrain on soil moisture retention, runoff, and vegetation growth. These factors collectively modulate the NDVI–WTD relationship within various geomorphological contexts. In regions of higher and middle relief, such as the Qilian Mountains, the correlations are strongly positive, suggesting that the environments in these areas support increased vegetation productivity. In contrast, lower relief regions, such as plains and platforms, exhibit a mixture of weak positive and negative correlations, indicating a more intricate interplay between groundwater availability, surface hydrology, and vegetation dynamics. Despite these variations, the majority of the correlation values in these lower relief areas remain positive, particularly in plains, platforms, and hills. With respect to the classes of water table depths (Figure 10b), the intermediate and relatively deep water table depths present the strongest NDVI–WTD correlations, whereas the shallow and deep regions present more varied correlation ranges. Notably, the deep water table class still demonstrates a substantial correlation, likely because vegetation that adapts to arid conditions relies on deep-rooted access to groundwater, often referred to as phreatophytes [9,45,47,100].

3.5. Association with Groundwater-Dependent Ecosystems (GDE) Area

On the basis of the above analysis, we know that the NDVI–WTD relationship is heavily influenced by both water availability and vegetation cover. For further investigation, we compared the areas with high NDVI–WTD correlations with a global groundwater-dependent ecosystem (GDE) map. Figure 11 maps the spatial distribution of the negative NDVI–WTD correlation patterns, as well as the GDE distribution. The spatial analysis reveals a notable overlap between regions with negative NDVI–WTD correlations (Figure 11a) and GDE distribution patterns (Figure 11b), with 32.30% of such areas aligning with mapped GDE zones. Within these overlapping regions, the distribution of GDE probabilities has a mean value of 0.588 (median = 0.530; standard deviation = 0.265), indicating a moderate to high level of confidence in groundwater dependence in these ecologically sensitive areas.

Importantly, however, the current GDE mapping algorithm excluded several parts of the HRB, especially the Qilian Mountain headwaters (uHRB) and alluvial fans (lHRB), which are critical zones for groundwater–vegetation interactions and where negative NDVI–WTD relationships are more likely to occur (see Figure S5). Therefore, the 32.30% overlap should be considered a conservative estimate, as the actual correlation between negative NDVI–WTD regions and GDE areas is likely greater than reported. This spatial alignment, even with incomplete GDE coverage, strongly supports the hypothesis that vegetation communities with negative NDVI–WTD correlations are disproportionately concentrated in GDEs. This is particularly important, as the exclusion of mountainous and alluvial fan regions—where groundwater–surface water interactions are the most dynamic—would likely enhance the observed correlation if these areas are incorporated into future GDE mapping efforts.

3.6. Contributions of Environmental Variables

The GAM is employed to quantify the contributions of various environmental factors influencing the NDVI–WTD relationship. The environmental variables are categorized into climate, soil, vegetation, hydrogeology, topography, and human activity. Prior to investigating the NDVI–WTD relationship using the GAM, a preliminary correlation analysis is conducted to assess the associations between the NDVI–WTD correlation and other environmental variables. The correlation matrix (Figure S6) shows the interdependencies among these variables. Notably, P exhibits the strongest correlation with the DEM (0.94). Additionally, P demonstrates a robust association with the SOM (0.72) and GPP (0.69), reinforcing its critical role in carbon cycling and ecosystem productivity. GSL exhibits strong positive correlations with the LAI (0.73), GPP (0.73), and PASW (0.72). The TI and vapor pressure deficit-related variables (LST, PAR, and CDD) show predominantly negative correlations with the soil moisture-related parameters (PASW and Ks). The highly negative correlation between the DEM and LST (−0.98) underscores the influence of elevation on the surface energy balance, with lower temperatures prevailing at higher elevations. In addition to these notable associations, most of the environmental variables exhibit relatively weak correlations, indicating a high degree of independence among them.

In Figure 12, we present a breakdown of the contributions of various factors influencing the NDVI–WTD relationships via the GAM, and these contributions are assessed using a combination of climate, soil, vegetation, hydrogeology, topography, and human activity. Overall, the environmental variables account for 87.9% of the spatial distribution of the NDVI–WTD correlation patterns. The analysis categorizes influential variables into six primary domains: climate, soil, vegetation, hydrogeology, topography, and human activity. The contributions of these factors are quantified and visualized in terms of the percentage of spatial variation that is explained. Climate-related variables contribute the greatest proportion of spatial variation, accounting for 26.59% of the total variance. Among these variables, the LST has the most substantial influence (11.9%), followed by P, PAR, and CDD. These factors play a significant role in controlling evapotranspiration and energy balance, thus affecting groundwater recharge and vegetation growth. The soil-related factors, including the PASW, SOM, pH, and ST, collectively account for 10.31% of the spatial variation. The role of vegetation, contributing 16.28%, is emphasized by metrics such as the MRD and GPP, which influence plant water use efficiency and carbon sequestration. Hydrogeological factors, particularly VZT and Ps, contribute 19.53% of the spatial variation, underscoring their role in subsurface water availability. Moreover, topographic parameters account for 7.05% of the variability, indicating the impact of elevation-driven hydrological processes. The anthropogenic factor (HM) contributes 8.14% of the overall variance, demonstrating the role of land-use change, infrastructure development, and agricultural management in modulating water availability and ecosystem responses. These percentages show the dominant role of climatic and hydrogeological factors in the NDVI–WTD relationships.

Figure 12b provides a detailed examination of the individual variables within each environmental category. Among the climate variables, the LST contributes the most, followed by the VZT and MRD. The influence of HM is also evident, as it ranks highly in its contribution. Soil and vegetation variables, such as SOM and LAI, have more moderate contributions, with the smallest impact observed from topographical variables such as TI and GSL. Finally, in Figure 12c, we summarize the total contributions by category. These results suggest that for effective collaborative vegetation–groundwater management, climate considerations, followed by vegetation and hydrogeological factors, should be prioritized, while also acknowledging the significant, though relatively small, influence of human activities and soil conditions.

4. Discussion

4.1. Application of the NDVI–WTD Relationships in Investigating Vegetation-Groundwater Interactions

Investigating vegetation–groundwater interactions through NDVI–WTD relationships is critical for quantifying ecosystem water dependencies, especially in arid regions where groundwater sustains both natural and managed vegetation under increasing hydrological stress. This approach enables spatially explicit monitoring of ecological sensitivity to groundwater fluctuations, informing sustainable water management and climate adaptation strategies. For example, D’Acunha et al. (2018) [42] linked rewetting success in disturbed peatlands to NDVI recovery, correlating rising water tables with Sphagnum recolonization through field-validated NDVI trends. Šimanauskienė et al. (2019) [26] established the NDVI as a cost-effective indicator of bog degradation in Lithuanian peatlands by demonstrating strong positive correlations between the NDVI and water table depth. Rohde et al. (2021) [51] revealed species-specific NDVI–water table relationships in California riparian woodlands, with deeper-rooted oaks showing stronger correlations than shallow-rooted willows. Xu et al. (2022) [33] mapped groundwater-dependent ecosystems in the Weihe River Basin via bivariate NDVI–WTD autocorrelation, highlighting synergies in mountainous regions with shallow aquifers. Zhao et al. (2022) [25] developed a regional-scale phreatic evapotranspiration model in arid Northwest China, validating the utility of the NDVI in resolving groundwater contributions to crop growth.

Our work in the arid HRB advances this field by demonstrating how NDVI–WTD relationships are modulated by hydroclimatic gradients and landcover heterogeneity—factors rarely addressed collectively in previous studies. The identification of contrasting NDVI–WTD correlations across hyper-arid versus arid zones, coupled with the quantification of climate as the dominant driver, provides a mechanistic framework for predicting the responses of vegetation to groundwater changes in water-limited systems. These findings complement global NDVI–WTD research by elucidating how human activities interact with natural drivers to shape vegetation–groundwater linkages in endorheic basins—a critical gap, given that arid zones worldwide rely on similar closed hydrological systems.

4.2. Utilization of Generalized Additive Models (GAMs) in Vegetation-Groundwater Studies

This study on spatiotemporal covariation in vegetation and groundwater dynamics within an endorheic inland river basin exemplifies this application, as the GAM framework enables the detection of subtle, smooth variations in the responses of vegetation that are driven by groundwater availability. Moreover, by incorporating penalization techniques to avoid overfitting, GAMs provide robust inferences, even when handling high-dimensional environmental datasets, making them ideally suited for advancing our understanding of ecosystem processes under variable climatic and hydrological conditions. This methodological approach not only facilitates the identification of key drivers of vegetation dynamics but also supports the development of more effective, data-informed water resource management strategies in arid and semiarid regions. In contrast to previous applications of GAMs in Earth sciences [56,57,98], in which a limited number of environmental factors have typically been considered, in this study, 19 variables related to climate, soil, vegetation, hydrogeology, topography, and human activity are incorporated. This comprehensive approach effectively demonstrates the capacity of the GAM to model complex, nonlinear environmental interactions.

4.3. Factors Influencing the Contributions of Environmental Variables

In this study, climatic (precipitation, temperature, radiation, and drought), soil (water retention, organic matter, pH, and texture), vegetation (root depth, productivity, and phenology), hydrogeological (hydraulic conductivity and aquifer properties), topographic (elevation and hydrological connectivity), and human activity (anthropogenic modification) variables are integrated to quantify their contributions to NDVI–WTD relationships. These factors collectively regulate water availability, subsurface dynamics, vegetation–climate feedback, and human-altered hydrological processes. The framework captures nonlinear interactions that drive vegetation–groundwater covariation across diverse arid ecosystems.

However, despite the comprehensive inclusion of these environmental variables, certain critical factors remain underrepresented. For example, soil classification, such as distinctions between clay-rich and sandy soils and their water retention capacities—both of which are essential for groundwater recharge efficiency and plant-available moisture—was not explicitly considered in the analysis. Furthermore, while human activity was broadly considered via the human modification index, a more granular approach to quantifying sector-specific water uses (such as agricultural irrigation withdrawals, industrial groundwater pumping, domestic consumption, and livestock demands) could provide a more mechanistic understanding of anthropogenic influences on the NDVI–WTD relationships. These limitations may reduce the predictive accuracy in scenarios where localized soil characteristics or differential human water use patterns play a dominant role in shaping vegetation–groundwater dynamics.

5. Conclusions

In this study, we investigate the spatiotemporal covariation between vegetation dynamics (NDVI) and groundwater table depth (WTD) in the Heihe River Basin (HRB), an arid inland basin characterized by hydroclimatic and geomorphological diversity. We first examine the seasonal dynamics of vegetation–groundwater interactions, followed by an evaluation of their spatiotemporal patterns using Spearman’s rank correlation and cross-correlation analyses. The variability in the NDVI–WTD relationships is then assessed across different landcover types and climate zones to determine their respective influences. Further investigations have explored the role of geomorphological characteristics in shaping vegetation–groundwater interactions, along with their associations with groundwater-dependent ecosystems (GDEs). Finally, the contributions of key environmental variables to the NDVI–WTD correlations are quantified via generalized additive models (GAMs) to capture nonlinear relationships and complex dependencies.

The seasonal dynamics of the NDVI and WTD in the HRB exhibit pronounced spatiotemporal heterogeneity, with summer characterized by a peak NDVI coinciding with shallow WTD and winter characterized by depressed vegetation activity alongside deeper water tables. The upstream regions demonstrate tight coupling between the NDVI and WTD fluctuations, whereas the midstream areas, with consistently deep groundwater, suggest vegetation reliance on mixed water sources. The downstream zones display minimal seasonal variability, which is indicative of surface water dependency. These patterns intensify during the growing season, particularly in groundwater-abundant areas, where hydroclimatic drivers strengthen the NDVI–WTD correlations. Spatial contrasts are the most evident in summer, when high water demand by vegetation aligns with shallow WTD, and in winter, where limited activity corresponds to deep groundwater. The transition seasons (spring/autumn) highlight shifting interacting regimes, with summer exhibiting the strongest NDVI–WTD synchrony.

The spatiotemporal analysis reveals a heterogeneous NDVI–WTD relationship across the HRB. Positive correlations dominate low-vegetation areas, whereas negative correlations prevail in regions with deeper water tables. Cross-correlation analysis reveals that 95% of the regions exhibit no time lag, indicating synchronous vegetation–groundwater relationships. Among lagged regions, short-term lags (1–3 months) are the most common. Regions with lagged responses are characterized by significantly greater soil water retention capacity (elevated PASW and SOM values) and constrained hydraulic connectivity (lower clay content, thinner aquifers, and reduced annual water table fluctuations).

Overall, groundwater depth adjustments largely follow vegetation changes within a short timeframe, highlighting the role of hydroclimatic and geomorphological factors in governing vegetation–groundwater interactions. Landcover–climate interactions significantly shape the NDVI–WTD correlations, with forests and grasslands in hyper-arid regions showing the strongest positive relationships (arid: 54.5%, hyper-arid: 28.9%), whereas croplands, shrublands, and arid wetlands present negative correlations (arid: 54.6%). Humid climates display weaker associations due to precipitation-driven groundwater stability, whereas synergistic vegetation–groundwater coupling emerges in dry sub-humid grasslands and arid wetlands. Spatial analysis reveals a 32.3% overlap between negative NDVI–WTD zones and mapped groundwater-dependent ecosystems (GDEs), with a mean GDE probability of 0.588 in overlapping areas. This alignment is likely underestimated, as current GDE mapping excludes critical HRB zones (e.g., Qilian headwaters, alluvial fans), suggesting stronger latent correlations between groundwater-dependent vegetation and negative NDVI–WTD dynamics. Finally, we determine the contributions of various environmental factors influencing the NDVI–WTD correlations via the GAM. Overall, the environmental variables account for 87.9% of the spatial distribution of the patterns of NDVI–WTD correlation. The largest contributing factor is climate, which accounts for 26.59% of the variance. This factor is followed by soil (10.31%), vegetation (16.28%), and hydrogeology (19.53%), whereas topography (7.05%) and human activity (8.14%) contribute significantly less.

The findings emphasize the need for integrated water resource management and land-use planning that considers both groundwater and vegetation dynamics. Model assumptions regarding groundwater–vegetation interactions align with observed real-world dynamics, particularly in regions where hydroclimatic drivers influence vegetation–water relationships. Future work should focus on exploring the temporal dimensions of these interactions, particularly in the context of climate and land-use changes, and assessing the potential impacts of future climate variability on groundwater-dependent vegetation systems. Furthermore, future studies should prioritize integrating spatiotemporal rainfall distribution plots at sub-basin scales to better characterize groundwater recharge mechanisms and their links to vegetation cover. Additionally, incorporating river discharge data could enhance the understanding of WTD fluctuations and improve model predictions under dynamic hydrological conditions.