Spatiotemporal Patterns of Soil Moisture Drought Across Different Soil Layers and Their Relationships with Ecosystem Water Use Efficiency and Resilience in the Three-North Shelterbelt Forest Program Region

Abstract

Ecosystem water use efficiency (WUE) is a key indicator of carbon–water coupling in ecosystems under climate change. While WUE responses to meteorological droughts are well-documented, the influence of different soil layers on WUE across vegetation types in water-limited regions remains unclear. Based on ERA5-Land soil moisture and MODIS vegetation products, this study investigates the spatiotemporal variations in soil moisture indices (SSMI) across three soil layers and their synchronous, lagged, and cumulative associations with WUE and resilience in the Three-North Shelterbelt Forest Program (TNSFP) region from 2001 to 2022. Our results showed that while the shallow and middle layers exhibited general wetting trends, the deep soil layer (100–289 cm) underwent continuous depletion in 56.01% of the study area. WUE showed weak synchronous responses to soil moisture drought across all layers, with no significant threshold effect. However, the lag and cumulative associations increased with soil depth, from 3–6 months in shallow layers to 9–12 months in deep layers, suggesting that deep-layer soil moisture may reflect longer-term ecohydrological stress and delayed ecosystem responses. Ecological resilience (R_d) differed by vegetation type. Grasslands were resilient to shallow drought but vulnerable to deep moisture deficits, while forests maintained stability under deep stress. These findings emphasize the importance of aligning vegetation configuration with soil water availability in restoration efforts, providing critical insights for sustainable water resource management and forest rehabilitation.

1. Introduction

Ecosystem water use efficiency (WUE), defined as the ratio of gross primary productivity (GPP) to evapotranspiration (ET), serves as a pivotal indicator of the coupling between carbon and water cycles in terrestrial ecosystems [1]. It reflects the trade-off between photosynthetic carbon uptake and water loss, thereby encapsulating the integrated responses of vegetation to environmental changes [2,3]. Under the background of global climate change, drought events have become more frequent and intense [4], posing severe threats to the stability of terrestrial ecosystems and the coupling of carbon–water cycles [5]. Drought directly alters soil moisture availability and atmospheric water demand, subsequently influencing plant physiological processes such as stomatal conductance, photosynthesis, and transpiration [6]. Therefore, understanding how WUE responds to drought is necessary for predicting ecosystem resilience, guiding sustainable water resource management, and supporting the sustainability of ecological restoration under changing climates.

Atmospheric vapor pressure deficit and soil moisture are extensively reported as the most significant environmental factors driving water movement and affecting photosynthesis and WUE [7,8]. Drought is a common hydrometeorological phenomenon [9], which can be broadly classified into meteorological drought, soil moisture drought, hydrological drought, and ecological drought [10]. In recent years, the responses of WUE to meteorological drought have been comprehensively explored. Yu et al. [11] integrated global-scale WUE data with the multi-temporal SPEI to systematically quantify the lagged and cumulative associations between drought and WUE. Liu et al. [12] utilized the SPEI at different scales to examine the temporal relationships between meteorological drought and WUE across three climate zones in the Yellow River Basin. Meanwhile, recent increasing evidence suggests that soil moisture drought is strongly associated with ecosystem productivity and water use [13,14]. Stocker et al. [15] found that a decline in soil moisture has the potential to reduce GPP by as much as 40% in water-constrained regions. In tropical rainforests, soil moisture drought shows a stronger association with ET than with photosynthesis, and WUE exhibits an increase under drought conditions [16]. However, most studies have focused on shallow soil moisture drought, while the comparison in dynamics of soil moisture drought across different soil layers and their long-term cumulative and lagged effects on ecosystem WUE remains unclear.

The influence of soil moisture on WUE may differ markedly between different soil layers [17]. Unlike shallow soil moisture, which fluctuates drastically with short-term precipitation pulses and evaporation, deep soil moisture maintains relatively stable content due to weak vertical migration and slow exchange with the atmosphere [18]. Vegetation with deep root systems, such as forests and woody shrubs, relies heavily on deep soil moisture to maintain photosynthetic capacity and transpiration balance [19]. Given the slower depletion and replenishment of deep soil moisture, drought signals in this layer may be associated with longer lagged and cumulative responses of WUE [20,21]. Furthermore, vegetation types differ in their stomatal regulation and carbon–water trade-off strategies [11]. Therefore, it is essential to systematically investigate the long-term cumulative and lagged associations between soil moisture drought and ecosystem WUE under different vegetation types.

The Three-North Shelterbelt Forest Program (TNSFP) represents one of the world’s largest ecological restoration projects, spanning arid, semi-arid, and semi-humid regions across northern China. As an important ecological barrier, the region’s ecosystem stability is closely related to regional water security and ecological balance [22]. Previous studies have shown that the TNSFP region has experienced significant vegetation greening in recent decades, which may be associated with changes in soil moisture consumption and ecosystem resilience [23]. However, how soil moisture drought across different layers is associated with variations in carbon–water cycle components (ET, GPP and WUE) through lagged and cumulative effects, and whether different vegetation types exhibit divergent resilience patterns are still lacking.

To address these knowledge gaps and provide scientific support for sustainable ecological restoration, this study explores the spatiotemporal patterns of soil moisture drought across three distinct layers, including shallow (0–28 cm), middle (28–100 cm), and deep (100–289 cm) layers in the TNSFP region from 2001 to 2022. The Standardized Soil Moisture Index (SSMI) was used to characterize soil moisture drought in this study. Compared with meteorological drought indices, SSMI directly reflects soil moisture anomalies relative to long-term normal conditions and can be calculated for different soil layers. Therefore, it is suitable for examining vertical differences in soil moisture drought and their relationships with ecosystem WUE. The specific objectives are to: (1) compare the relationships between WUE and soil moisture drought across multiple soil layers; (2) identify the soil layer most strongly associated with WUE in different vegetation types; (3) characterize the spatiotemporal patterns, lagged and cumulative associations of drought in these layers; and (4) assess the resilience of ecosystems to soil moisture drought stresses. To direct these objectives, we formulated four testable hypotheses: (1) soil moisture drought trends differ among shallow, middle, and deep soil layers; (2) the relationship between WUE and SSMI becomes more delayed and cumulative with increasing soil depth; (3) woody vegetation may show longer lagged and cumulative associations with deep soil moisture drought than herbaceous vegetation; (4) ecosystem resilience may differ among vegetation types and drought severities.

2. Materials and Methods

2.1. Study Area

The TNSFP region encompasses 13 provincial-level administrative divisions in the northeastern, northern, and northwestern parts of China (33°30′ N–50°12′ N, 73°26′ E–127°50′ E), covering approximately 4,069,000 km². The terrain in this region decreases from west to east with relatively complex topography. Most of the region has an annual mean air temperature ranging from 2 °C to 11 °C and annual precipitation ranging from 50 mm to 800 mm. It is characterized by an arid and semi-arid monsoon climate, limited water resources, and frequent sand and dust storms [24]. Since the project was launched, good results have been achieved in the construction of forest and grassland vegetation and the control of desertification. According to the MODIS Land Cover Type Product (MCD12Q1), the vegetated area accounts for 59.392% of the TNSFP region (Figure 1). The study area consists of various predominant vegetation types, including needleleaf forests (NF, 0.384%), broadleaf forests (BF, 4.672%), mixed forests (MF, 0.461%), shrublands (SHR, 0.078%), savannas (SAV, 4.013%), grasslands (GRA, 36.073%), and croplands (CRO, 13.711%).

2.2. Data and Processing

2.2.1. Soil Moisture Drought Index

Soil moisture data from 2001 to 2022 were obtained from the ERA5-Land reanalysis dataset (https://apps.ecmwf.int/datasets/, accessed on 10 December 2025). The ERA5-Land soil moisture data have a spatial resolution of 0.1° and include four soil layers: 0–7 cm, 7–28 cm, 28–100 cm, and 100–289 cm. Daily data were averaged to derive monthly soil moisture. ERA5-Land was selected because it provides long-term, spatially continuous, and vertically stratified soil moisture data, which are suitable for analyzing soil moisture drought across different soil layers at regional scale [25]. ERA5-Land dataset has been widely used to investigate soil water dynamics across shallow, intermediate, and deep soil layers, and its reliability in China has been supported by previous studies [18]. In the dry zone of Central Asia, these datasets were validated against ISMN in situ observations, showing a significant relationship between ERA5-Land soil moisture and field measurements [19].

This study divided the soil profiles into three layers to investigate soil moisture droughts, including the shallow soil layer (0–28 cm), the middle soil layer (28–100 cm), and the deep soil layer (100–289 cm). The original ERA5-Land 0–7 cm and 7–28 cm layers were combined into the shallow layer (0–28 cm) using the thickness-weighted averaging method. The Standardized Soil Moisture Index (SSMI) provides valuable information about soil moisture droughts that might not be apparent through other drought indices [26]. The corresponding SSMI values were denoted as SSMI-1 for 0–28 cm, SSMI-2 for 28–100 cm, and SSMI-3 for 100–289 cm, and were calculated as follows:

$$SSMI = {\displaystyle \frac{{SM}_{ij}-{\overline{SM}}_i}{\sigma }}$$

(1)

where SM_ij is the monthly soil moisture value for the ith month in j year, and SM_i is the mean soil moisture value for the ith month of the multiyear period from 2001 to 2022. According to previous studies [27], SSMI < 0 was used in this study to identify below-normal soil moisture conditions.

2.2.2. Vegetation Characteristics

Vegetation water use efficiency (WUE) is defined as the ratio of gross primary productivity (GPP) to actual evapotranspiration (ET). The monthly NDVI dataset from 2001 to 2022 was derived from the MOD13A3 Version 6.1 product. Monthly GPP and ET were generated from the 8-day composite MOD17A2H Version 6.1 and MOD16A2 Version 6.1 products, respectively, both of which are available at a 500 m spatial resolution. MOD17A2H provides 8-day gross primary productivity, while MOD16A2 provides 8-day actual evapotranspiration estimates based on the Penman-Monteith framework. To obtain monthly totals, the 8-day composite values overlapping with a given month were first converted to daily rates and then weighed by the exact number of calendar days falling within that month. This procedure ensured temporal consistency with the monthly SSMI data and the subsequent monthly lagged and cumulative analyses. ET was retained and reported as a water-depth variable in mm. For the calculation of WUE, 1 mm ET was treated as equivalent to 1 kg H₂O m⁻²; therefore, WUE was calculated as the ratio of monthly GPP to monthly ET and expressed as g C kg⁻¹ H₂O⁻¹. It should be noted that WUE is a derived indicator rather than an independently observed variable. Both MODIS-derived GPP and ET are algorithm-based products and may be affected by meteorological inputs, vegetation parameters, land-cover classification, and model assumptions. Therefore, uncertainties in GPP and ET may propagate into MODIS-derived WUE, especially in areas with low ET. These datasets are available from the US Geological Survey (https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 15 December 2025).

All datasets were first projected to the WGS84 geographic coordinate system and clipped to the boundary of the TNSFP region. To ensure spatial consistency, ERA5-Land soil moisture data were resampled to 0.05° using bilinear interpolation to match the spatial resolution of the MODIS-derived vegetation datasets. For categorical land-cover data, the nearest-neighbor method was used. Non-vegetated areas, water bodies, urban lands, barren lands, and permanent snow/ice were masked using the MCD12Q1 Version 6.1 land-cover product. Only vegetated pixels, including needleleaf forests, broadleaf forests, mixed forests, shrublands, savannas, grasslands, and croplands, were retained for subsequent analysis.

MODIS quality assurance layers were used to remove low-quality observations. Missing values caused by cloud contamination or low-quality retrievals were filled only when the missing period was shorter than two consecutive months using linear interpolation; otherwise, the pixel was excluded from the corresponding temporal analysis. All variables were aggregated or resampled to monthly time steps from 2001 to 2022. GPP and ET were converted to consistent monthly units before calculating WUE.

2.3. Methods

2.3.1. Trend Analysis

To quantify temporal trends in SSMI, NDVI, GPP, and WUE from 2001 to 2022, the Theil-Sen slope method is employed in our study. The Theil-Sen slope method computes the median of all possible pairwise slopes between distinct time steps using the formula:

$$\mathit{\beta } = \mathrm{median}\left({\displaystyle \frac{x_j-x_i}{j-i}}\right) (j > i)$$

(2)

where β is the slope of the variables, j and i represent time steps in the time series, xj and xi represent the variable values at the time steps j and i, respectively. Subsequently, the Mann–Kendall (MK) test is applied to evaluate the statistical significance of the trends estimated by the Theil-Sen slope. The MK test quantifies the rank correlation between the time series values and their corresponding time indices. By applying the Theil-Sen slope and MK test at the pixel scale, we also identify the spatial heterogeneities in trend patterns across vegetation types. The involved formulas are calculated as follows:

$$S={\displaystyle {\sum }_{i=1}^{n-1}}{\displaystyle {\sum }_{j=i+1}^n}sgn\left(x_j-x_i\right)$$

(3)

where sgn(x_j − x_i) equals 1, 0, or −1 when (x_j − x_i) is greater than, equal to, or less than zero, respectively. The statistic S is standardized into Z, and trends are considered significant when $\left|Z\right|>$ 1.96 at the 0.05 significance level.

2.3.2. Lagged and Cumulative Associations Between Drought and Vegetation

Before correlation analyses, all monthly variables, including SSMI, NDVI, GPP, ET, and WUE, were converted into anomalies to remove seasonal cycles. For each pixel and each variable, the long-term mean of each calendar month during 2001–2022 was subtracted from the original monthly value. The anomaly series were then detrended using linear regression to reduce the influence of long-term co-trending patterns. The lagged and cumulative correlation analyses were conducted using the deseasonalized and detrended anomaly series.

The normality of the deseasonalized and detrended anomaly series was examined using the Shapiro–Wilk test for regional and vegetation-type time series. Because many monthly anomaly series did not strictly satisfy the normality assumption, Spearman rank correlation was used to quantify the relationships between SSMI and vegetation indicators. To reduce the influence of temporal autocorrelation, the effective degrees of freedom were estimated according to the lag-1 autocorrelation coefficients of the two time series, and the significance of correlation coefficients was tested using the adjusted effective sample size. To reduce the risk of inflated Type I error caused by repeated scanning, false discovery rate (FDR) correction was applied separately across the 0–12-month lag windows and the 1–12-month cumulative windows. The R_maxlag and R_maxcum were extracted only from time windows that passed both the autocorrelation-adjusted significance test and FDR correction; otherwise, the corresponding pixels were treated as non-significant.

The lagged association between soil moisture drought and vegetation indicators was characterized by Spearman rank correlation analysis. Spearman rank correlation analyses were conducted between monthly NDVI, GPP, and WUE data and the SSMI from 1- to 12-month scales. Among the significant lag windows, the correlation coefficient with the maximum absolute value was selected as the maximum lagged correlation coefficient (R_maxlag). The involved formulas were calculated as follows:

$$R_i = corr(VI, {SSMI}_i) \mathrm{with} p_{\mathrm{FDR}} < 0.05, 1 \le i \le 12$$

(4)

$$R_{maxlag}=R_k \mathrm{where} \left|R_k\right|=\underset{\mathrm{i}\in \Omega }{\max }\left(\left|R_i\right|\right)$$

(5)

where VI is the monthly vegetation indices of NDVI, GPP, and WUE, and Ri is the Spearman correlation coefficient between vegetation and drought with a lagged time of i months. Ω represents the subset of lag months (1 to 12 months) that successfully passed both the temporal autocorrelation-adjusted significance test and the FDR correction (pFDR < 0.05). Among this significant subset Ω, the specific month k exhibiting the peak absolute correlation magnitude ($\left|R_k\right|$) was identified. R_maxlag retained its original directional sign to preserve its distinct ecohydrological implications.

For the cumulative association analysis, Spearman rank correlation was conducted between monthly vegetation indicators and cumulative SSMI values across 1–12-month accumulation windows. The j-month cumulative SSMI was calculated as the arithmetic mean of the SSMI from the current month back to month j. Consistent with the lagged analysis, the correlation coefficients (Rj) were strictly regulated by temporal autocorrelation adjustments and throttled via FDR correction (p_FDR < 0.05) across the 1–12-month cumulative windows. The greatest correlation coefficient was identified as the maximum cumulative correlation coefficient (R_maxcum). The involved equations were calculated as follows:

$$R_j = corr(VI, {SSMI}_j) \mathrm{with} p_{\mathrm{FDR}}< 0.05, 1 \le j \le 12$$

(6)

$$R_{maxcum}={ R}_m \mathrm{where} \left|R_m\right|=\underset{\mathrm{j}\in \Lambda }{\max }\left(\left|R_j\right|\right)$$

(7)

Here, R_j was the correlation coefficient between vegetation and drought at accumulated j months. Λ denoted the specific subset of cumulative months (1 to 12 months) that exhibited statistically significant relationships after passing the dual criteria of autocorrelation correction and FDR thresholding (p_FDR < 0.05). R_maxcum was identified as the native coefficient at month m that achieved the absolute maximum value within the set Λ, keeping its original positive or negative directional sign intact to accurately reflect the directional feedback of cumulative drought stress.

2.3.3. Ecosystem Resilience Analysis

According to Sharma and Goyal [28], the ratio of the mean WUE during drought periods to that during non-drought periods, presented by the index of Rd, was employed to evaluate the ecosystem resilience. This study uses this index to quantify the resilience of ecosystems across different vegetation types to deep soil moisture drought. The following equation is used:

$$R_d = {\displaystyle \frac{{WUE}_d}{{WUE}_m}}$$

(8)

where Rd is the WUE-based ecosystem resilience index, WUEd is the mean value of the WUE during soil moisture deficit periods (SSMI ≤ 0), and WUEm is the mean value of the WUE during non-deficit periods (SSMI > 0). The value of Rd can be classified into four categories: (1) resilient (Rd ≥ 1); (2) slightly non-resilient (1 > Rd ≥ 0.9); (3) moderately non-resilient (0.9 > Rd ≥ 0.8); (4) severely non-resilient (Rd < 0.8).

2.3.4. Uncertainty Quantification

To quantify the uncertainty of the main statistical results, we quantified the uncertainty by calculating the 95% confidence intervals (CIs) for all primary trend slopes, correlation coefficients, and resilience indices. The CIs for Theil-Sen trend slopes (β) were derived using the non-parametric method, ensuring the statistical reliability of the temporal trends in SSMI, GPP, and WUE. For the ecological resilience index (Rd), we implemented a bootstrapping approach with 1000 iterations to estimate the 95% CIs for each vegetation type, providing a rigorous assessment of the stability patterns under drought stress.

3. Results

3.1. Variations in Soil Moisture Drought and Ecosystem Carbon–Water Cycle Components

3.1.1. Spatiotemporal Coupling and Divergence of Soil Moisture Drought Across Layers

The soil moisture droughts occurred uniformly among seasons (Figure 2). The soil moisture drought in the shallow soil layer (0–28 cm) persisted throughout 2001–2022, whereas the drought in the middle layer (28–100 cm) and deep layer (100–289 cm) were predominantly observed after 2005 and 2007, respectively. The monthly SSMI-1 of 2007–2008, SSMI-2 of 2006–2009 and 2014–2015, and SSMI-3 of 2008–2012, 2015–2018, and 2020 were less than 0, indicating the soil moisture deficit duration tended to be longer with increasing soil depth. As shown in Figure 3a, the mean annual SSMI-1, SSMI-2 and SSMI-3 during 2001–2022 exhibited significant spatial variability across the TNSFP region, with an average of 0.00 ± 0.03, 0.00 ± 0.05 and −0.04 ± 0.09, respectively. Overall, SSMI-1 and SSMI-2 displayed upward trends, with positive pixels constituting 51.28% and 55.91%, respectively, predominantly concentrated in the northeastern region (Figure 3b). Conversely, the SSMI-3 showed a downward trend, with negative pixels making up 56.01%, distributed across the northern part of the Loess Plateau, the Jungger Basin, and the northeastern area.

3.1.2. Spatial–Temporal Variations in Annual ET, GPP, and WUE

The spatial distribution of the mean annual ET, GPP, and WUE during the period of 2001–2022 in the TNSFP region was shown in Figure 4a,c,e. The spatial distribution of mean annual ET and GPP exhibited similar patterns, with greater values distributed in southeastern areas and lower values in northern and western areas. The WUE across the region showed spatial variability, ranging from 0 to 4.02 g C kg⁻¹ H₂O⁻¹ with a mean value of 1.1 g C kg⁻¹ H₂O⁻¹. In the northwestern and northern regions, the WUE appeared to be inversely correlated with the distribution of ET and GPP, whereas WUE showed a similar distribution pattern to ET and GPP in the southeastern region. From 2001 to 2022, the vegetation ET, GPP, and WUE exhibited upward trends with changing rate of 4.28 ± 0.45 mm yr⁻¹ (95% CI: 3.83–4.73), 6.01 ± 0.62 g C m⁻² yr⁻¹ (95% CI: 5.39–6.63), and 0.0028 ± 0.0003 g C kg⁻¹ H₂O⁻¹ yr⁻¹ (95% CI: 0.0025–0.0031) (Figure 4a). As shown in Figure 4b,d, the area proportions showing significant increasing trends of ET and GPP were 77.4% and 77.5% (p < 0.05), respectively, while 0.5% and 0.4% of the area experienced significant decreasing trends. The WUE showed increasing trends in 58.0% of the area, with 19.4% in a significant upward trend (p < 0.05), distributed primarily in central areas.

Among different vegetation types, forests including NF, BF, and MF demonstrated the greatest WUE of 1.80, 1.79, and 1.76 g C kg⁻¹ H₂O⁻¹, respectively, whereas GRA showed the lowest WUE of 1.12 g C kg⁻¹ H₂O⁻¹ (Figure 5b). The SHR showed the greatest increasing rates in ET and GPP at 5.32 mm yr⁻¹ and 10.20 gC m⁻² yr⁻¹, respectively. The CRO exhibited the lowest increasing rates in WUE at 0.0024 g C kg⁻¹ H₂O⁻¹ yr⁻¹. The NF had the lowest increasing rates in ET (0.94 mm yr⁻¹) and GPP(4.29 gC m⁻² yr⁻¹), but the greatest rate in WUE (0.0077 g C kg⁻¹ H₂O⁻¹ yr⁻¹).

3.2. Divergent Responses of Ecosystem WUE to Soil Moisture Droughts

Figure 6 shows that the fitted relationships between WUE at the 1-month scale and SSMI-1, SSMI-2, and SSMI-3 were all statistically non-significant (p > 0.05). Specifically, WUE exhibited weak positive linear relationships with SSMI-1 and SSMI-2, whereas a weak quadratic relationship was observed for SSMI-3. Compared with WUE, ET and GPP exhibited relatively clearer but still non-significant responses to soil moisture drought. ET generally showed a decreasing tendency with increasing drought severity, and this decline became more evident under deeper soil drought conditions. In contrast, GPP varied only slightly under shallow and middle soil drought, but tended to decrease under deep soil drought.

Moreover, the responses of WUE to soil moisture drought varied across different vegetation types (Figure A1 and Figure A2). However, these fitted relationships were generally weak and showed large scatter, suggesting that short-term soil moisture is not the sole driver. Since the synchronous analysis only captures the immediate response of WUE, whereas ecosystem WUE may show delayed and cumulative associations with drought conditions, further analyses were conducted to examine the temporal associations between WUE and soil moisture drought across different soil layers.

3.3. Lagged and Cumulative Associations Between Soil Moisture Drought and WUE

3.3.1. Lagged Associations Between Soil Moisture Drought and WUE

As shown in Figure 7 and Figure 8, the lagged associations between soil moisture drought and ecosystem WUE exhibited pronounced spatial heterogeneity and clear differences in dominant lagged months across the three soil layers in the TNSFP region. For SSMI-1, lagged associations were mainly distributed in the central-eastern and parts of the northeastern region, where negative correlations were more extensive than positive ones (Figure 7a). The corresponding lagged months were mainly concentrated in short- to medium-term scales, with relatively shorter lagged durations in the central-southern areas and longer lagged durations occurring locally in the northeastern region (Figure 7b). Positive lagged associations under SSMI-1 peaked at 3 months (Figure 8a), whereas negative lagged associations reached the greatest proportion at 6 months, indicating that WUE was primarily negatively associated with shallow soil moisture drought at medium-term lagged scales.

For SSMI-2, the spatial pattern of R_{max_lag} became more heterogeneous, with positive and negative correlations interwoven across the central-eastern areas (Figure 8c). Meanwhile, the lagged months expanded from short-term to medium- and long-term scales, especially in the northeastern and eastern parts of the study area (Figure 8d). Figure 8b showed that positive lagged associations were most evident at 9 months, whereas negative associations were relatively weak and dispersed among different lagged months. These results suggested that the associations between WUE and middle-layer soil moisture drought were more delayed and temporally dispersed than those observed for shallow-layer soil moisture drought.

For SSMI-3, positive lagged associations became more extensive than in the shallower layers, particularly in the northeastern and eastern regions, whereas negative associations were mainly distributed in the central and southwestern margins (Figure 7e). At the same time, the corresponding lagged months shifted further toward medium- and long-term scales (Figure 7f). Positive lagged associations became markedly stronger at 9–12 months, while negative associations were mainly concentrated at 6–8 months (Figure 8c). These patterns indicate that deep-layer soil moisture drought was associated with a stronger delayed legacy signal than shallow- and middle-layer soil moisture drought.

Under the different vegetation types, the NF, BF, and MF exhibited longer lagged durations than others for the SSMI at all three soil layers, while GRA and CRO showed the shortest duration (Figure 8d,f). As shown in

Table A1. Area percentages of positive and negative R_{max_lag} and lagged time scales under different vegetation types.

Vegetation Types	Type of SSMI	Area Percentage (%)
		Rmax_Lag		Lagged Months
		Positive	Negative	Short Term	Medium Term	Long Term
				(1–4 Months)	(5–8 Months)	(9–12 Months)
NF	SSMI-1	10.028	89.972	19.767	78.682	1.550
	SSMI-2	63.225	36.775	35.112	34.730	30.158
	SSMI-3	32.033	67.967	21.846	48.923	29.231
BF	SSMI-1	25.977	74.023	34.318	65.617	0.065
	SSMI-2	30.016	69.984	34.673	47.528	17.799
	SSMI-3	44.155	55.845	23.645	44.819	31.536
MF	SSMI-1	27.430	72.570	30.139	69.453	0.408
	SSMI-2	34.118	65.882	25.272	37.012	37.716
	SSMI-3	49.808	50.192	24.301	48.534	27.164
SHR	SSMI-1	6.793	93.207	88.449	11.394	0.157
	SSMI-2	16.564	83.436	37.683	31.126	31.190
	SSMI-3	18.750	81.250	51.940	15.476	32.584
SAV	SSMI-1	21.347	78.653	40.838	57.630	1.533
	SSMI-2	26.821	73.179	37.683	31.126	31.190
	SSMI-3	56.293	43.707	31.002	22.116	46.882
GRA	SSMI-1	21.107	78.893	44.388	42.288	13.324
	SSMI-2	39.808	60.192	26.834	54.816	18.351
	SSMI-3	75.952	24.048	15.388	23.693	60.919
CRO	SSMI-1	26.435	73.565	31.939	60.705	7.356
	SSMI-2	29.490	70.510	54.153	34.235	11.612
	SSMI-3	87.143	12.857	14.900	11.379	73.721

, all vegetation types for SSMI-1 were dominated by negative lagged associations, with negative area proportions ranging from 72.570% in MF to 93.207% in SHR. For SSMI-2, NF was the only vegetation type dominated by positive lagged associations, while the other vegetation types still showed larger negative lagged areas. For SSMI-3, positive lagged associations exceeded negative ones in SAV, GRA, and CRO.

3.3.2. Cumulative Associations Between Soil Moisture Drought and WUE

For SSMI-1, negative cumulative correlations were widely distributed in the central-eastern areas, whereas positive cumulative correlations only occurred in scattered patches, mainly in the northeastern and marginal areas (Figure 9a). The corresponding cumulative months were predominantly concentrated at short-term scales, especially in the central and southwestern parts of the region (Figure 9d). Figure 10a showed that negative cumulative associations reached the greatest area proportion at the 1-month scale, while positive cumulative associations mainly peaked at 4 months.

For SSMI-2, the spatial pattern of R_{max_cum} became more heterogeneous (Figure 9c). The cumulative months expanded from short-term dominance to a mixture of short-, medium-, and long-term scales, especially in the northeastern and eastern regions (Figure 9d). Correspondingly, Figure 10b showed that positive cumulative associations became more evident at 3 and 12 months, whereas negative associations were mainly concentrated at 1, 7, and 8 months.

For SSMI-3, positive cumulative associations became more extensive than in the shallower layers, particularly in the central-western and northern parts of the TNSFP region, whereas negative cumulative associations were mainly retained in localized areas of the northeast and southwest (Figure 9e). At the same time, the corresponding cumulative months shifted further toward medium- and long-term scales, indicating a longer-term cumulative association between deep soil drought and WUE (Figure 9f). Figure 10c showed that positive cumulative associations became markedly stronger at 1, 11, and 12 months, while negative cumulative associations were relatively weaker and mainly concentrated at short cumulative scales.

Similar to the lagged duration among the vegetation types, the NF, BF, and MF showed a longer lagged duration than others for SSMI-1, SSMI-2, and SSMI-3, while GRA and CRO presented a shorter duration (Figure 10d,f). All vegetation types for SSMI-1 were dominated by negative cumulative associations (

Table A2. Area percentages of positive and negative R_{max_cum} and cumulative time scales under different vegetation types.

Vegetation Types	Type of SSMI	Area Percentage (%)
		Rmax_Cum		Cumulative Months
		Positive	Negative	Short Term	Medium Term	Long Term
				(1–4 Months)	(5–8 Months)	(9–12 Months)
NF	SSMI-1	7.521	92.479	73.538	19.499	6.964
	SSMI-2	55.292	44.708	52.939	34.433	12.628
	SSMI-3	40.947	59.053	69.638	19.499	10.864
BF	SSMI-1	36.641	63.359	68.467	30.913	0.620
	SSMI-2	52.416	47.584	51.968	34.475	13.557
	SSMI-3	44.332	55.668	55.736	30.636	13.628
MF	SSMI-1	21.077	78.923	74.846	23.103	2.051
	SSMI-2	40.136	59.864	56.403	28.699	14.898
	SSMI-3	53.243	46.757	52.059	40.479	7.462
SHR	SSMI-1	3.321	96.679	93.297	5.888	0.815
	SSMI-2	36.814	63.186	59.905	21.065	19.030
	SSMI-3	16.033	83.967	91.214	0.483	8.303
SAV	SSMI-1	26.614	73.386	79.595	19.070	1.335
	SSMI-2	46.178	53.822	64.401	13.129	22.469
	SSMI-3	45.543	54.457	59.801	19.253	20.946
GRA	SSMI-1	14.570	85.430	81.809	16.695	1.496
	SSMI-2	49.982	50.018	60.052	20.645	19.303
	SSMI-3	74.501	25.499	51.918	9.868	38.215
CRO	SSMI-1	45.775	54.225	50.608	46.543	2.849
	SSMI-2	48.379	51.621	64.082	27.637	8.280
	SSMI-3	77.099	22.901	69.320	7.634	23.046

). For SSMI-2, cumulative responses became more balanced between positive and negative signs. For SSMI-3, positive cumulative associations became dominant in MF, GRA, and CRO, accounting for 53.243%, 74.501%, and 77.099% of area with significant correlations, respectively.

3.4. Ecosystem Resilience Under Soil Drought Stresses

The spatial distribution of ecosystem resilience under vertical soil moisture deficit conditions is presented in Figure 11. In general, resilient and slightly non-resilient ecosystems dominated the TNSFP region under all three soil layers, while moderately and severely non-resilient areas were relatively limited and mainly distributed in scattered patches of the central and western parts of the study region. For SSMI-1, slightly non-resilient areas were widely distributed in the central-eastern areas, whereas resilient areas mainly appeared in the northeastern and northwestern margins (Figure 11a). For SSMI-2, resilient areas became more widespread and continuous than under SSMI-1, especially in the northeastern region (Figure 11b). For SSMI-3, the resilient areas further expanded in the northeastern and northwestern parts of the TNSFP region, while moderately non-resilient areas remained mainly scattered in the central transition zone (Figure 11c). These spatial patterns indicate that ecosystem resilience generally exhibited stronger expression under middle-layer and deep-layer soil moisture deficit than under shallow-layer soil moisture deficit in many vegetated areas.

Table 1. Ecosystem resilience analysis for different vegetation types.

Vegetation Types	Type of SSMI	Lowest-SSMI Year	Mean SSMI for the Lowest-SSMI Year	WUEd	WUEm	Rd (Mean ± 95% CI)	Ecosystem Resilience
NF	SSMI-1	2007	−0.843	1.814	1.906	0.951 (0.942–0.960)	Slight non-resilience
	SSMI-2	2008	−1.267	1.866	1.885	0.990 (0.981–0.999)	Slight non-resilience
	SSMI-3	2008	−1.506	1.918	1.884	1.018 (1.002–1.034)	Resilience
BF	SSMI-1	2008	−0.523	1.780	1.863	0.955 (0.945–0.965)	Slight non-resilience
	SSMI-2	2008	−0.731	1.840	1.857	0.991 (0.982–1.000)	Slight non-resilience
	SSMI-3	2008	−0.957	1.850	1.860	0.995 (0.986–1.004)	Slight non-resilience
MF	SSMI-1	2007	−0.614	1.785	1.844	0.968 (0.958–0.978)	Slight non-resilience
	SSMI-2	2008	−0.829	1.834	1.827	1.004 (0.958–0.978)	Resilience
	SSMI-3	2008	−1.152	1.846	1.833	1.007 (0.998–1.016)	Resilience
SHR	SSMI-1	2009	−0.534	1.852	1.841	1.006 (0.996–1.016)	Resilience
	SSMI-2	2007	−0.584	1.560	1.540	1.013 (1.003–1.023)	Resilience
	SSMI-3	2010	−0.943	1.822	1.856	0.982 (0.973–0.991)	Slight non-resilience
SAV	SSMI-1	2007	−0.938	1.677	1.740	0.964 (0.954–0.974)	Slight non-resilience
	SSMI-2	2008	−1.066	1.722	1.712	1.006 (0.996–1.016)	Resilience
	SSMI-3	2008	−1.354	1.743	1.728	1.008 (0.998–1.018)	Resilience
GRA	SSMI-1	2006	−0.387	1.570	1.526	1.029 (1.019–1.039)	Resilience
	SSMI-2	2007	−0.382	1.482	1.517	0.977 (0.967–0.987)	Slight non-resilience
	SSMI-3	2011	−0.337	1.513	1.541	0.982 (0.972–0.992)	Slight non-resilience
CRO	SSMI-1	2008	−0.522	1.599	1.619	0.988 (0.978–0.998)	Slight non-resilience
	SSMI-2	2008	−0.758	1.115	1.141	0.992 (0.982–1.002)	Slight non-resilience
	SSMI-3	2009	−0.905	1.613	1.621	0.995 (0.985–1.005)	Slight non-resilience

Note: WUE_d and WUE_m indicate mean WUE during soil moisture deficit periods and non-deficit periods, respectively. R_d = WUE_d/WUE_m. The 95% CI of R_d was estimated using bootstrap resampling with 1000 iterations.

further showed that the resilience of different vegetation types varied distinctly with soil depth. For SSMI-1, only SHR and GRA were classified as resilient, with Rd values of 1.006 and 1.029, respectively, while the other vegetation types were slightly non-resilient. Among them, NF showed the lowest R_d value (0.951), followed by BF (0.955), SAV (0.964), MF (0.968), and CRO (0.988). For SSMI-2, the resilience pattern shifted toward woody vegetation. MF, SHR, and SAV were classified as resilient, with Rd values of 1.004, 1.013, and 1.006, respectively, whereas NF, BF, GRA, and CRO remained slightly non-resilient. SHR exhibited the strongest resilience under SSMI-2, while GRA showed the weakest resilience with an Rd value of 0.977. For SSMI-3, NF, MF, and SAV were resilient, with Rd values of 1.018, 1.007, and 1.008, respectively, whereas BF, SHR, GRA, and CRO were slightly non-resilient. NF showed the highest resilience under deep-layer soil moisture deficit conditions, while SHR and GRA both declined to an Rd value of 0.982. Overall, the resilience pattern indicated that the shallow-rooted vegetation such as GRA showed stronger resilience under shallow soil drought, whereas woody vegetation such as NF, MF, and SAV exhibited stronger resilience under middle-layer and deep-layer soil drought conditions.

4. Discussion

4.1. Spatiotemporal Coupling Relationship Between Soil Drought and Ecosystem WUE

This study revealed that ecosystem WUE exhibits pronounced spatial heterogeneity in response to soil moisture drought in the TNSFP region, with notable differences across the three soil layers, including 0–28 cm (SSMI-1), 28–100 cm (SSMI-2), and 100–289 cm (SSMI-3). Shallow soil moisture (0–28 cm) is highly sensitive to short-term precipitation and evaporation, leading to frequent fluctuations (Figure 2). In contrast, the middle (28–100 cm) and deep soil layers (100–289 cm) generally show slower temporal variations than shallow soil moisture. Therefore, they may provide information on longer-term soil water availability during dry periods [18,19]. While SSMI-1 and SSMI-2 generally showed upward trends with positive pixels covering 51.28% and 55.91% of the area, SSMI-3 exhibited a distinct downward trend with negative pixels making up 56.01% of the region. It was suggested that while shallow and middle soil layers may be recovering due to precipitation or management, the deep soil layer is experiencing a continuous deficit.

The spatial distribution shows that higher soil moisture levels are concentrated in central areas, while lower values are found in the northern and western arid zones. These patterns are closely linked to the different sensitivities of GPP and ET to water shortages across the region [29]. In the arid western parts of the TNSFP region, vegetation exhibits strong adaptation to drought conditions. These plants respond quickly to water limits by reducing transpiration (ET) more significantly than photosynthetic capacity (GPP), which leads to an increase in WUE [6]. In the central region where deep soil moisture is persistently depleted, the buffering effect of subsoil water weakens, and prolonged water limitation may eventually constrain both GPP and ET, thereby reducing the stability of WUE [30]. This suggests that the sustainability of vegetation restoration in these areas depends not only on increases in vegetation productivity or coverage, but also on whether deep soil water reserves can support long-term carbon–water coupling. In contrast, dense forests in the southeastern region utilize stable deep soil moisture to support high levels of both transpiration and carbon fixation, resulting in higher observed WUE [14].

Consistent with the global pattern of WUE enhancement [11,13], the ET, GPP, and WUE showed increasing trends in the TNSFP region from 2001 to 2022. Such asynchronous changes in ET and GPP may offset each other in the calculation of WUE, thereby weakening the overall synchronous response of WUE to soil moisture drought. In this study, ET generally declined with increasing drought severity, and the decline became more evident with increasing soil depth, whereas GPP remained relatively stable under shallow-layer and middle-layer soil moisture deficit but decreased more clearly under deep-layer soil drought. This suggested that under shallow drought, reduced water loss may partly offset the impact of drought on carbon uptake, thereby weakening the immediate response of WUE. By contrast, under deep drought, prolonged water limitation may constrain photosynthetic carbon assimilation more strongly, leading to a different WUE response pattern [30]. Figure 3 suggests that differences in WUE among diverse vegetation types were closely related to their utilization capacity of deep soil moisture. The higher observed WUE in forest ecosystems may be partly related to their greater capacity to access deep soil water, which is consistent with their deeper rooting systems [31], maintaining stable GPP during droughts. In contrast, croplands and grasslands with shallow roots rely more on shallow soil moisture [32], resulting in lower WUE and weaker coupling with deep soil moisture.

The response of WUE to soil moisture drought in the TNSFP region differs from the threshold effect reported by Zhang et al. [20], who found that the WUE of most vegetation types peaked at moderate soil drought (SSMI = −0.5). However, our study revealed that the relationships between WUE and SSMI were nonlinear and non-significant and also did not detect a threshold. This discrepancy may be attributed to the distinct characteristics of vegetation adaptation strategies [33]. The shallow soil moisture fluctuates drastically with short-term precipitation and evaporation, leading to a clear threshold in the trade-off between GPP and ET [34]. In contrast, the deep soil layer maintains relatively stable moisture content [18] and serves as a buffering water reservoir for vegetation.

Unlike some previous studies that identified a threshold effect where WUE peaks at moderate drought levels [12], our study found that the synchronous responses of WUE to soil moisture were weak and non-significant across all three soil layers (Figure 6). No clear threshold effect was detected for SSMI-1, SSMI-2, and SSMI-3. This lack of a threshold response can be attributed to the asynchronous changes in ET and GPP under drought stress within the TNSFP region. As drought severity increases, the simultaneous decline in ET and the relative stability of GPP may offset each other in the WUE calculation, thereby weakening the immediate coupling between WUE and soil moisture status at the monthly scale [19]. This suggests that the TNSFP ecosystems may rely more on longer-term cumulative and lagged adjustments rather than immediate threshold-based physiological responses to manage water stress.

4.2. Lagged and Cumulative Associations Between Soil Moisture Drought and WUE

Previous studies have provided evidence that ecosystem WUE not only responded instantaneously to meteorological and soil moisture drought but also exhibited a lagged and cumulative response following the occurrence of droughts [20]. This study systematically examined these effects across three distinct soil layers, revealing a clear vertical shift in the timing of drought associations. In the shallow soil layer (0–28 cm), the lagged associations were primarily negative and concentrated in short- to medium-term scales, typically peaking at 3–6 months. These findings aligned with previous studies that suggested that shallow moisture deficits quickly force plants to adjust their water use strategies to prevent dehydration, primarily through stomatal closure to reduce water loss, thereby increasing WUE in the short term [35]. However, this adaptive response is often temporary and can be overwhelmed when drought conditions persist for longer periods [36]. As drought stress moves into the middle (28–100 cm) and deep (100–289 cm) layers, the lagged duration significantly increases, with impacts often reaching 9–12 months. This delayed transmission of drought signals may be related to the slower response of deep soil moisture to short-term climate fluctuations [18]. Unlike shallow soil moisture, which is more responsive to rapid changes in precipitation and evaporation, the deep soil moisture exhibits a more gradual depletion, with longer-term physiological effects on vegetation, including root development and stomatal regulation [33].

The negative lagged associations are dominant in the shallow layer, indicating that drought is often associated with a temporary increase in WUE as plants close their stomata to reduce water loss [7]. As shown in Figure 2, our study area predominantly experienced light and moderate drought (−1 < SSMI < 0). Under light and moderate drought, plants responded to drought via stomatal closure to reduce water loss at the expense of carbon assimilation and consequently increased WUE and negative lagged associations [35]. Moreover, abundant precipitation during the growing season was likely to offset legacy effects of previous droughts [37]. However, positive lagged associations became more widespread in the deep soil layer. These positive associations suggested that deep moisture deficits may be associated with a decline in WUE by overcoming the adaptive limits of the ecosystem [5]. This finding highlighted the widespread sensitivity of even shallow-rooted vegetation to deep soil moisture, especially when surface layers became dry. It also underscored the potential importance of deep soil moisture for WUE stability during prolonged drought conditions, particularly for vegetation types with deep root systems [17].

The cumulative association analysis further confirms that most vegetated areas experienced cumulative associations, and deeper-layer soil moisture deficit showed a more persistent and longer-lasting association with WUE than shallow-layer soil moisture deficit (Figure 7). Shallow soil drought (SSMI-1) generally was associated with WUE through short-term cumulative responses (1–4 months), which was consistent with that in the Yellow River Basin (4.50 months) [12], and the duration was longer than that in the Ziya River Basin (2.11 months) [14]. In the TNSFP and Yellow River Basins, which were characterized by arid and semi-arid conditions, moderate water limitations and dominance of soil moisture drought led to relatively longer cumulative associations. In contrast, the Ziya River Basin, with extensive cropland and intensive irrigation, exhibited shorter cumulative durations, which may be related to continuous moisture replenishment from irrigation, which interrupted deep soil moisture depletion [38]. In the TNSFP region, deeper soil drought influenced WUE over much longer cumulative time scales, where long-term cumulative associations accounted for a large portion of the affected area. This suggested that deep soil moisture depletion may be associated with persistent water stress and longer-term changes in the balance between GPP and ET [39]. Such long-lasting cumulative associations imply that sustainable vegetation management in the TNSFP region should account for delayed ecosystem responses rather than relying solely on short-term drought indicators.

Among different vegetation types in the TNSFP region, forests (NF, BF and MF) exhibited the longer average lagged and cumulative duration, while CRO had the shorter duration (Figure 7 and Figure 9). This pattern was closely related to the structural and functional characteristics of vegetation. Forests with deep and extensive root systems can access deep soil moisture to buffer immediate drought stress [31], leading to a delayed lagged response. Meanwhile, their high-water demand and long growth cycle mean that continuous deep soil moisture deficit accumulates over time, resulting in longer cumulative duration [40]. The short lagged and cumulative durations of CRO were primarily due to anthropogenic management (e.g., irrigation) that supplements soil moisture, reducing the persistence of drought stress [41]. Additionally, crops had a short growth cycle and rapid physiological turnover, leading to faster responses to drought and weaker cumulative associations compared to natural vegetation [32].

4.3. Vertical Variations in Ecological Resilience to Soil Moisture Drought

Ecological resilience to soil drought refers to the ability of ecosystems to maintain or recover their structure and function under water stress, which was a key indicator of ecosystem stability [42]. Our analysis shows that while resilient and slightly non-resilient areas dominate the region, the specific ability of different vegetation types to maintain WUE stability depends on the depth at which drought occurs. Under shallow soil moisture drought (SSMI-1), only SHR and GRA were classified as resilient, with Rd values of 1.006 and 1.029, respectively (Table 1). This high resilience in shallow layers is likely due to the rapid physiological adjustment of shallow-rooted species, which allows them to effectively capture pulse precipitation and recover quickly from short-term moisture deficits [19]. In contrast, forest ecosystems showed lower resilience in the shallow layer, with needleleaf forests exhibiting the lowest R_d value of 0.951. As drought stress moves into the middle and deep soil layers, the pattern of resilience shifts significantly in favor of woody vegetation. For deep soil drought, NF, MF, and SAV were classified as resilient, with NF reaching its highest R_d value of 1.018. This higher resilience in deeper layers may be partly related to the greater ability of woody plants to access stable subsoil moisture through deeper rooting systems [21].

This vertical shift suggested that the long-term stability of the TNSFP region depends on the ability of its diverse vegetation types to utilize water from different soil depths based on their specific rooting strategies. However, the resilience of GRA and SHR declined to 0.982 under deep-layer soil moisture deficit conditions. This decline supports the growing evidence that even though grasslands are primarily shallow-rooted, their ecosystem functions are increasingly sensitive to the status of deep soil moisture reserves [17]. When deep soil water is exhausted due to prolonged drought or excessive consumption by surrounding forests, the protective buffering capacity for grasslands is lost, leading to a state of non-resilience [43]. These findings have significant implications for the sustainable management of the TNSFP project. The vertical bifurcation of resilience emphasizes the necessity of matching vegetation types with the water availability of specific soil layers [6]. To prevent the unsustainable depletion of deep soil water, which we identified as a growing concern in the Loess Plateau and other key regions, forest management must consider stand characteristics such as planting density and species composition [44]. Promoting a balanced distribution of shallow-rooted and deep-rooted species can enhance the overall ecological resilience of the region to multi-layer soil moisture drought. Such strategies are essential for improving the sustainability of large-scale ecological restoration in water-limited regions.

4.4. Limitations

Although this study provided a comprehensive perspective on the vertical responses of WUE to soil moisture drought, several limitations should be acknowledged. First, WUE responses to soil moisture drought may be confounded by atmospheric and climatic drivers. In northern China, strengthened land–atmosphere coupling and high VPD may intensify soil moisture stress by promoting earlier stomatal closure and altering both GPP and ET [6]. Therefore, future investigations should combine soil moisture, atmospheric aridity, radiation, flux observations, and vegetation structural information to better disentangle their individual and interactive effects on ecosystem carbon–water coupling.

Second, although SSMI is useful for representing soil moisture anomalies across different soil layers, it depends on the accuracy of the soil moisture dataset. ERA5-Land provides continuous multi-layer soil moisture information, but its deep-layer estimates (100–289 cm) are model-derived and less directly constrained by satellite and in situ observations than surface soil moisture [19]. GLDAS/Noah, GLEAM root-zone soil moisture, and GRACE-based terrestrial water storage can provide useful references, but they are not directly equivalent to ERA5-Land deep-layer soil moisture because they differ in soil-depth definition, water-storage components, model structure, and spatial resolution [25]. Therefore, the deep-layer depletion identified in this study should be interpreted as reanalysis-based evidence and requires further validation using in situ deep soil moisture observations.

5. Conclusions

This study evaluated the spatiotemporal evolution of multi-layer soil moisture drought and its associations with ecosystem WUE and resilience in the TNSFP region from 2001 to 2022. The main conclusions are as follows: (1) while the shallow (0 to 28 cm) and middle (28 to 100 cm) soil layers generally showed upward trends, the deep soil moisture (100 to 289 cm) exhibited a continuous deficit in 56.01% of the study area, particularly in the northern Loess Plateau and the Jungger Basin; (2) the synchronous relationship between ecosystem WUE and soil moisture drought was generally weak across all three soil layers, and no significant threshold effects were detected. However, the dominant lagged and cumulative associations shifted toward longer time scales with increasing soil depth, from 3 to 6 months in the shallow layer to 9–12 months in the deep layer. These results suggest that deep-layer soil moisture was associated with longer-term variations in WUE and may provide information on delayed ecohydrological responses; (3) ecosystem resilience showed vertical differences among vegetation types. Grasslands showed higher resilience under surface moisture fluctuations but lower resilience under deep-layer soil moisture deficits. Conversely, forest ecosystems showed higher stability under deep-layer stress, which may be partly related to their deeper rooting systems and greater access to subsoil water. These findings suggest that restoration management should optimize stand structure and species composition according to soil water availability across different soil layers for ecosystem sustainability.