Impact Pathways of Environmental Factors on the Spatiotemporal Variations in Surface Soil Moisture in Tianshan Mountains, China

Abstract

Soil moisture (SM) in the mountains is critical for agropastoral productivity, and it is subject to both large-scale climate gradients and fine-scale effects of terrain, vegetation and soil. However, how the climate, topography, soil and vegetation factors impact surface SM spatiotemporal dynamics remains elusive in mountainous terrains, due to their complex interactions. Based on multi-source datasets, this study employs the structural equation model to investigate the impact pathways of climate and vegetation factors on annual surface SM dynamics from the year 2000 to 2022 in the Tianshan Mountains of China (TS). We also utilize the factor and interaction detectors of Geographical Detector to explore the individual and interactive effects of climate, topography, soil and vegetation factors on the spatial pattern of the annual surface SM. Moreover, their integrated impacts on the spatiotemporal dynamics of annual surface SM were investigated based on the explanatory power from the factor detector and total effects from structural equation modeling. The results showed that the multi-year average surface SM was 0.21 m³·m⁻³ for the whole region, with greater values in areas with dense vegetation and high elevation. Annual surface SM exhibited significant increasing trends across different land cover classifications and elevation zones, which was directly influenced by vegetation greenness enhancement. Precipitation (PRE) and relative humidity (RH) also significantly influenced the temporal variations in surface SM through their indirect effect on vegetation greenness, while these indirect effects were much lower than the direct effect of vegetation greenness. RH, PRE and surface net solar radiation (SSR) showed strong individual and interactive effects on the spatial distribution of surface SM, particularly the interactive effects of RH and PRE with wind speed (WS). Surface SM was highly sensitive to RH and PRE in the central TS. Overall, vegetation greenness, PRE and RH were the main drivers of surface SM variations across both temporal and spatial scales, while SSR, total evaporation and WS primarily shaped its spatial distribution. These insights enhance our understanding of land–atmosphere interactions in mountainous areas and provide scientific references for sustainable agropastoral water resource management under global warming.

1. Introduction

Soil moisture (SM) is a crucial variable of agropastoral systems, as it directly determines water availability for vegetation growth and influences water use efficiency and thus affects agropastoral productivity and sustainable water resource management [1,2]. For example, SM regulates the partitioning of precipitation into infiltration and runoff, thereby controlling water availability, water uptake, and water stress during critical vegetation growth stages, especially under dry conditions [3]. Meanwhile, climate factors, such as precipitation and air temperature, are widely acknowledged as the primary drivers of the spatiotemporal variations of SM [4,5,6]. Moreover, topography, soil properties, land cover classifications (LCCs) and vegetation features also play significant roles in shaping SM distribution characteristics [7,8]. For example, precipitation and vegetation cover dominated SM dynamics in grassland, while it was precipitation in bare land [7]. Moreover, soil organic carbon can affect the soil bulk density, which then alters the soil water retention and volume [8].

Despite extensive research on SM variability across diverse environments, our understanding of mountainous SM remains incomplete and contested. In flat or gently undulating landscapes, climatic controls on SM have been well documented [9]. However, in complex mountain terrain, the interplay among atmospheric drivers, topography, and subsurface characteristics produces scale-dependent, heterogeneous SM behavior. For example, although the relationship between SM and topography varies by location in high-latitude regions [10], topographical factors associated with condensation and evapotranspiration processes were critical to SM in the Anning Basin, located in Southwest China [11]. Recent work indicates that SM memory and its dominant controls shift across temporal scales in mountain watersheds, with dynamic transitions from event-driven to long-term storage regulation that are not fully understood [12]. Moreover, fine-scale hillslope processes such as lateral subsurface flow and slope-aspect effects on rainfall response challenge the assumption of spatially uniform controls [13]. These findings suggest differing views on the relative roles of climate, topography, and soil properties in mountainous SM dynamics, indicating the need for more integrative approaches that reconcile the temporal and spatial mechanisms of SM.

Due to the sparse distribution of in situ SM and meteorological stations in mountainous and desert regions [14], gridded datasets are often required for regional analyses. Among satellite-derived products, the European Space Agency’s Climate Change Initiative SM dataset, providing a long-term record since 1978, exhibits substantial data gaps in topographically complex regions. In high-elevation areas, retrieval uncertainties caused by terrain effects, snow cover, frozen soil, and heterogeneous vegetation further reduce data availability, affecting spatial continuity and reliability [15]. In contrast, the Soil Moisture Active Passive mission provides relatively high-accuracy L-band SM retrievals; however, its short temporal coverage (since 2015) and data screening under frozen or snow-covered conditions limit its use for long-term analyses [16]. The Global Land Evaporation Amsterdam Model (GLEAM), a process-based model driven by satellite and reanalysis data, estimates SM based on water balance and energy constraints. By explicitly representing evaporation, transpiration, and SM dynamics, it has been shown to perform reliably across arid and high-elevation ecosystems, making it particularly suitable for characterizing moisture variability in mountainous, arid, and semi-arid regions [17,18,19].

Previously, multiple linear regression, correlation analysis, and partial correlation analysis have been widely employed to investigate the relationships between SM and its influencing factors, such as climate, vegetation, and topography [20,21,22,23]. However, these methods may not capture the complex interactions and impact pathways of the drivers. Partial Least Squares Structural Equation Modeling (PLS-SEM) is a variance-based structural equation modeling technique that combines the advantages of path analysis and factor analysis, enabling the examination of direct, indirect, and total effects among variables [24,25]. PLS-SEM places minimal constraints on measurement scales, sample sizes, and residual distributions, making it particularly suitable for revealing latent structures and implicit relationships within observed data [26,27]. It has been widely applied in disciplines such as social sciences, ecology, and hydrology to explore the relationships among multiple causes and outcomes within complex systems [28,29,30,31]. Meanwhile, Geographical Detector (GeoDetector) evaluates the relationship between independent and dependent variables through spatial distribution similarity [32]. Compared to traditional linear regression models, the GeoDetector model utilizing statistical methods not only handles both categorical and numerical variables to explore dominant factors but also quantifies the interaction between two variables without requiring linear assumptions for the dependent variable [33]. This method has been widely applied to analyze the extent to which the independent variables influence the spatial heterogeneity of the dependent variable [34,35,36]. Recently, studies in mountainous and arid regions have successfully combined GeoDetector and PLS-SEM to investigate complex environmental dynamics. For example, Zhang et al. (2025) applied this framework to the spatiotemporal dynamics of flash floods in the Tianshan Mountains, and Cao et al. (2025) employed the same approach to disentangle the differential impacts of natural and anthropogenic factors on ecological environmental quality across human and non-human activity zones on the northern slope of the Tianshan Mountains [37,38]. Therefore, to investigate the complex interactions and mechanisms shaping SM, this study integrates PLS-SEM and GeoDetector to identify the dominant drivers and their underlying pathways.

The Tianshan Mountains (TS), known as the Water Tower of Central Asia [39], are part of the world’s high-elevation regions. Influenced by the westerlies and topographic barriers, the TS exhibit a distinct continental climate, characterized by a gradual increase in temperature and precipitation from the southeast to the northwest, as well as significant variations in temperature and humidity across different elevations [40]. Recent research indicates that high-latitude regions have experienced persistent warming over the past century [41], and the TS have likewise undergone significant climatic changes characterized by increasing temperature [40], alongside a reduction in terrestrial water storage, and a decline in groundwater levels [42,43]. Due to limited water availability and high vegetation water consumption, the TS frequently experience SM deficits [44]. Research on SM in the TS has mainly focused on its temporal variations and the corresponding impacts from hydrothermal variables [20,45]. However, previous studies have rarely quantified the relative importance versus interaction effects of multiple environmental drivers, nor have they explicitly examined pathway-based mechanisms underlying SM variability. Moreover, an integrated long-term spatiotemporal assessment that simultaneously considers the spatial heterogeneity and temporal dynamics of surface SM remains lacking in the TS.

Although irrigation, grazing, and land management may influence SM locally, this study focuses on regional-scale variability driven by natural environmental factors. Anthropogenic effects are not explicitly included, and the results should be interpreted within this context. Owing to insufficient environmental and SM observations in the TS, this study utilized reanalysis meteorological data, reanalysis SM data, satellite remote sensing elevation and land cover classification data to identify the impact pathways and intensities of environmental factors on the spatiotemporal dynamics of surface SM. The main objectives of this study are (1) to identify the spatial distribution characteristics of surface SM and its environmental factors from 2000 to 2022, along with their temporal dynamics during this period; (2) to elucidate the impact pathways and magnitudes of meteorological and vegetation factors on the temporal dynamics of surface SM by PLS-SEM; (3) to identify both the individual and interactive effects of environmental drivers on the spatial heterogeneity of surface SM by GeoDetector; and (4) to explore the integrated effects of environmental factors on the spatiotemporal dynamics.

Based on these objectives, we formulated the following research hypotheses: H1: Surface soil moisture (SM) shows significant spatial heterogeneity across different elevation gradients and land cover classifications, accompanied by corresponding differences in the dominant environmental drivers. H2: The strength and pathways of environmental controls on SM vary across elevation gradients and land cover classifications, resulting in distinct temporal and spatial influence patterns.

The findings will deepen our understanding about the mechanism of various environmental factors on SM in complex topography, offering scientific insights into land–atmosphere interactions and supporting agropastoral water resource management.

2. Materials and Methods

2.1. Study Area

The Tianshan Mountains are one of the largest mountain systems in Asia, characterized by an overall east–west orientation. About 90% of the TS is located within China, spanning from 38°56′ to 45°34′ N and 72°45′ to 95°53′ E (Figure 1). Located in arid central Asia, the annual average temperature and precipitation in the TS are about 6.8 °C and 218 mm, respectively. In the past few decades, precipitation in the TS has generally increased, accompanied by a rise in temperature, especially in high-elevation areas [46,47]. Due to their unique geographical location, the TS with vulnerable ecosystems are sensitive to climate change [48,49]. The average elevation of the study area is 2050 m above sea level (a.s.l.), with the highest point exceeding 7000 m a.s.l. (Figure 1a). The central and western parts show high elevations, while the eastern parts are relatively low. The main land cover classifications in 2020 (Figure 1b) included bare land (with an area ratio of 48.78%), grassland (32.86%), cropland (12.63%), water bodies (0.76%), forest (2.37%), permanent snow and ice (2.54%) and urban areas (0.06%). The topsoil organic carbon content in the study area is high in the central and northern regions (Figure 1c), with an overall average of 11.1 g·kg⁻¹, while the topsoil bulk density is low in the central and northern regions but high in the eastern and southern regions, with an overall average of 1.45 g·cm⁻³. The topsoil organic carbon content in the study area is high in the central and northern regions (Figure 1c). The topsoil bulk density is low in the central and northern regions, but high in the eastern and southern regions (Figure 1d). Soil texture exhibits a clear west–east gradient, with heavier, clay-rich soils in the west favoring organic carbon accumulation, and looser soils in the east leading to greater nutrient loss [50].

2.2. Data Sources

To investigate the factors driving the spatiotemporal dynamics of surface SM in the TS from 2000 to 2022, we selected the normalized difference vegetation index (NDVI), precipitation (PRE), surface net solar radiation (SSR), land surface temperature (T), wind speed (WS), relative humidity (RH), total evaporation (TE, where negative values indicate evaporation and positive values indicate condensation), Slope, Elevation (Elev.), Aspect, land cover classifications (LCCs), topsoil organic carbon (SOC) and topsoil bulk density (BULK) as potential contributors to SM spatiotemporal variability. The corresponding datasets used in this study and their details are summarized in

Table 1. Datasets used in this study.

Variable	Dataset	Spatial Resolution	Temporal Resolution	Source	References
Surface soil moisture (SM)	GLEAM	0.25° × 0.25°	Monthly	https://www.gleam.eu/ accessed on 27 November 2023	[51]
Surface net solar radiation (SSR), Total evaporation (TE)	ERA5-Land	0.1° × 0.1°	Monthly	https://cds.climate.copernicus.eu/ (accessed on 14 November 2023)	[52]
Relative humidity (RH)	ERA5	0.25° × 0.25°	Monthly	https://cds.climate.copernicus.eu/ accessed on 10 November 2024	[53]
Near surface wind speed (WS)	FLDAS	0.1° × 0.1°	Monthly	https://disc.gsfc.nasa.gov/ (accessed on 26 July 2023)	[54]
Land surface temperature (T)	TRIMS LST	1 km × 1 km	Daily	https://data.tpdc.ac.cn/ (accessed on 11 August 2023)	[55]
Precipitation (PRE)	1 km monthly precipitation dataset for China	1 km × 1 km	Monthly	https://data.tpdc.ac.cn/ (accessed on 6 December 2023)	[56]
NDVI	MOD13A3 v6.1	1 km × 1 km	Monthly	https://ladsweb.modaps.eosdis.nasa.gov/ (accessed on 24 October 2023)	[57]
Elevation (Elev.), Aspect, Slope	ASTER GDEM	30 m × 30 m	-	https://www.gscloud.cn/ (accessed on 30 August 2023)	[58]
Land cover classification (LCC)	Land cover classification gridded maps from 1992 to present derived from satellite observations	300 m × 300 m	Yearly	https://cds.climate.copernicus.eu/ (accessed on 24 January 2025)	[59]
Soil bulk density (BULK), Soil organic carbon content (SOC)	HWSD	1 km × 1 km	-	https://www.fao.org/home/en/ (accessed on 17 January 2024)	[60]

.

The surface SM data used in this study were obtained from the Global Land Evaporation Amsterdam Model (GLEAM) reanalysis dataset [51]. Although the applicability and accuracy of the GLEAM SM data have been widely demonstrated in previous studies [17,18,61], in situ hourly measured surface SM data at 5 cm depth, obtained from the Tianshan Snow Cover and Avalanche Research Station from April 2018 to March 2019, were further used to validate its performance in this study. The results show that the correlation coefficient between GLEAM and the in situ surface SM data was significant at 1%, with an RMSE of 0.037 m³·m⁻³ and an MAE of 0.027 m³·m⁻³.

The ERA5-Land Monthly Averaged ECMWF Climate Reanalysis dataset is derived from the land component of the ECMWF ERA5 climate reanalysis [52]. Both the ERA5 and ERA5-Land datasets have been widely validated for their reliability and applicability [53,62,63]. SSR and TE were obtained from the ERA5-Land dataset, while RH was obtained from ERA5.

The FLDAS Noah Land Surface Model is a system that combines observational, remote sensing, and forecast data to generate high-quality climate variables for monitoring, agricultural forecasting, and disaster risk assessment [54,64], and WS from the FLDAS Noah LSM was used in this study.

The daily 1 km all-weather land surface temperature dataset for China’s landmass and its surrounding areas (TRIMS LST) is generated using the enhanced Reanalysis and Thermal Infrared Remote Sensing Merging (E-RTM) method, which combines Terra/Aqua MODIS LST products and GLDAS data, along with ancillary data such as the NDVI and surface albedo. The dataset has been validated against MODIS LST and 19 in situ stations, showing high spatial consistency and low biases [55]. It was used as the T dataset in this study.

The 1 km monthly precipitation dataset for China provides monthly precipitation data at a spatial resolution of 0.0083°. Generated in China using the Delta spatial downscaling method, this dataset is based on global 0.5° data from CRU and high-resolution data from WorldClim. The dataset was validated against 496 independent meteorological stations, confirming its accuracy and reliability [56]. It was used as the PRE dataset in this study.

The MODIS Vegetation Indices Monthly (MOD13A3 v6.1) data, provided by the Terra satellite, are available as a Level 3 product with a 1 km spatial resolution and provide monthly NDVI data since February 2000 [57]. In this study, the NDVI from MOD13A3 was used as vegetation greenness. It has been widely adopted in vegetation monitoring and agricultural studies for quantifying vegetation dynamics and distinguishing vegetation types [65].

ASTER GDEM (Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model) is a global digital elevation model derived from data collected by NASA’s Terra satellite. With a 30 m resolution, it covers the Earth’s surface from 83° N to 83° S latitude [58]. Elev., Slope, and Aspect were extracted from ASTER GDEM in this study, with Elev. further classified into four categories based on terrain conditions, namely, Low (<1000 m), Medium (1000–3500 m), High (3500–5000 m), and Very High (>5000 m) [66], to facilitate the investigation of SM variation mechanisms across different elevation gradients. The Very-High-elevation zone was excluded from the SM analysis due to permafrost and glaciers there, which prevent the formation of significant soil layers [67].

The LCC gridded maps from 1992 to present provide land cover data for the study area with an original resolution of 300 m [59]. Twenty-two land use types were reclassified into seven categories: cropland, forest, grassland, bare land, water bodies, permanent snow and ice, and urban areas. Water bodies, permanent snow and ice, and urban areas were excluded from the SM analysis, as these land cover types lack active soil layers or meaningful SM dynamics. Such exclusions are common practice in large-scale SM studies to ensure reliable and interpretable results [68].

Soil properties data were sourced from the Harmonized World Soil Database (HWSD version 1.2), developed by the Food and Agriculture Organization (FAO) and the International Institute for Applied Systems Analysis (IIASA) [60]. Soil properties, including soil bulk density and soil organic carbon content, significantly influence the distribution of SM [69]. Therefore, top soil bulk density (BULK) and topsoil organic carbon content (SOC) were selected as drivers of surface SM in this study. Soil properties, as well as terrain variables, were treated as time-invariant due to the lack of consistent long-term dynamic datasets at the regional scale, which may overlook gradual changes such as land degradation, soil compaction, or erosion that could alter SM retention over time.

The meteorological and land surface variables were mainly obtained from established reanalysis and remote sensing products (e.g., ERA5, ERA5-Land, GLEAM, FLDAS, MOD13A3 v6.1, ASTER GDEM and HWSD) with inherent quality control. For the TRIMS LST and the 1 km monthly precipitation dataset for China, extreme outliers were removed using physical thresholds, and missing values were filled via inverse distance weighting, ensuring spatiotemporal consistency across datasets. The meteorological, vegetation greenness (NDVI) and LCC datasets used in this study cover the period from 2000 to 2022. To focus on the interannual responses of SM to environmental factors in high-mountain regions, the original monthly or daily data were aggregated to annual mean or cumulative values, thereby reducing seasonal fluctuations and facilitating the analysis of long-term trends and potential causal relationships. Specifically, SM, T, WS, SSR, NDVI, and RH were represented by annual mean values, whereas TE and PRE were represented by annual cumulative values. Elev., Slope, Aspect, SOC, and BULK were considered to remain unchanged throughout the study period. Resampling was performed to ensure consistency among different data sources in subsequent analyses and to eliminate biases caused by differences in resolution, allowing the datasets to be more effectively compared and integrated. All datasets used in this study were resampled to 0.0083° resolution, with LCCs resampled using the majority resampling method and the remaining datasets resampled using the bilinear interpolation. The geographic coordinate system for all datasets was transformed to WGS1984 to ensure spatial consistency.

2.3. Methods

The research framework and analytical workflow of this study is shown in Figure 2.

2.3.1. Quantitative Assessment of Resampling Uncertainty

To evaluate the potential spatial uncertainty introduced by resampling various datasets to a 0.0083° grid, we conducted a spatial variance preservation analysis to quantitatively assess the retention of spatial heterogeneity.

Firstly, for the variables that required resampling, we calculated the coefficient of variation ratio (CV) and the variance ratio (Var) for both the original and resampled datasets. These metrics quantify the spatial dispersion and pattern characteristics of the data.

Secondly, we defined the ratio for the coefficient of variation (CV_ratio) and the variance ratio (Var_ratio) as follows:

$$\begin{array}{l}CV\_ratio={\displaystyle \frac{C{V}_{resampled}}{C{V}_{original}}}\\ Var\_ratio={\displaystyle \frac{Va{r}_{resampled}}{Va{r}_{original}}},\end{array}$$

(1)

These two metrics quantify the extent to which resampling preserves spatial heterogeneity. Var_ratio < 1 suggests that resampling introduces smoothing by reducing spatial variance, while Var_ratio ≈ 1 indicates minimal impact of resampling on spatial variance. Similarly, CV_ratio ≈ 1 implies that resampling has little effect on the coefficient of variation, thus maintaining the spatial dispersion of the data [70].

The assessment showed that resampling high-resolution topographic and land cover data largely preserved their spatial characteristics. Specifically, for LCCs, the Var_ratio was 0.95 and the CV_ratio was 1.03, while for DEM, the Var_ratio was 1.01 and the CV_ratio was 1.00. Similarly, resampling of multi-year averaged low-resolution environmental variables largely preserved their spatial characteristics. Specifically, for SM, SSR, TE, RH, and WS, the Var_ratio values were 0.78, 0.97, 0.88, 0.91, and 0.85, respectively, and the CV_ratio values were 1.12, 1.01, 0.93, 0.94, and 0.88, respectively. Overall, although resampling may induce some local smoothing, the primary spatial patterns and heterogeneity of the datasets are largely preserved at the 0.0083° analysis resolution, supporting the robustness of subsequent multivariate spatial analyses.

2.3.2. Trend Analysis

To determine the long-term trends and magnitudes of surface SM and its major driving factors during 2000–2022, this study combined the Mann–Kendall (M–K) trend test with the Theil–Sen median slope estimator. The M–K test is a non-parametric approach used to assess the statistical significance of monotonic trends in time series and is insensitive to data distribution and outliers. The Theil–Sen method estimates the rate of change by calculating the median of pairwise slopes, providing a robust measure of trend magnitude [71,72]. A significance level of p < 0.05 was adopted to identify statistically significant trends. The integration of these two methods enables simultaneous evaluation of trend significance and change rate, thereby revealing the temporal evolution characteristics of SM and its driving factors over the study period. Detailed procedures and formulas can be found in [38].

2.3.3. PLS-SEM

PLS-SEM was used to reveal the direct effects (represented by standardized path coefficients), indirect effects (calculated through mediator variables) and composite total effects of each factor influencing annual surface SM dynamics. The applicability of the PLS-SEM model was assessed using non-parametric evaluation metrics, including the statistical significance of path coefficients (p-values), averaged coefficient of determination for all latent variables (R²), Variance Inflation Factor (VIF) from collinearity statistics, Standardized Root Mean Square Residual (SRMR), and Normed Fit Index (NFI) [73]. PLS-SEM results were acceptable and reliable when meeting the following thresholds: p-value < 0.05, average R² > 0.33, VIF < 3, SRMR < 0.05, and NFI > 0.90 [74].

Using 23 years of annual surface SM data, along with seven indicating factors (PRE, SSR, T, WS, RH, TE, and NDVI), we investigated their effects on the temporal variations in surface SM across three dimensions by PLS-SEM: (1) the entire study area, (2) three elevation gradients (low, medium and high), and (3) four land cover classifications (cropland, forest, grassland, and bare land).

Based on the minimum sample size estimation procedures proposed by Kock and Hadaya (2018) [26], the inverse square root method was applied to determine whether the available observations are sufficient for PLS-SEM analysis. Assuming a significance level of 5% and a statistical power of 80%, the minimum required sample size is calculated as:

$$n_{\min }>{\left({\displaystyle \frac{2.486}{p_{\min }}}\right)}^2,$$

(2)

where p_min represents the minimum expected path coefficient considered substantively meaningful. In environmental process studies, only moderate-to-strong structural relationships are typically interpreted as practically meaningful. Adopting a conservative minimum expected effect size of 0.50–0.60, the required minimum sample size is approximately 17–25. The available 23 observations fall within this estimated range, indicating that the sample size is acceptable for detecting moderate-to-strong structural effects in PLS-SEM. For each PLS-SEM model, the maximum number of structural paths pointing to a single endogenous variable (SM) is four. Based on 23 annual observations, the ratio of the sample size to the number of structural parameters is approximately 5.75:1. This indicates a moderate sample-to-model complexity relationship, suggesting that the available observations are sufficient for estimating the specified PLS-SEM models.

When PLS-SEM was constructed by SmartPLS (v4.1.1.1) software, we followed a structured procedure: (1) the full model was run including all indicator factors; (2) whether the model met the following criteria was evaluated: VIF < 3, p-value < 0.05, and R² > 0.33; (3) explanatory variables with non-significant paths (p-value ≥ 0.05) or multicollinearity issues (VIF ≥ 3) were removed stepwise according to the established thresholds; the theoretical causal structure of the model remained unchanged throughout the optimization process, and only statistically unsupported paths were removed; (4) the refined model was re-estimated and further evaluated using VIF, p-value, R², SRMR < 0.08, and NFI > 0.90; and (5) the final optimized model was obtained, from which the direct, indirect, and total effects of the retained variables were subsequently estimated.

To assess the statistical significance and robustness of the estimates, a bootstrapping procedure with 5000 subsamples was performed, using a two-tailed test at a significance level of 0.05 and a fixed random seed to ensure reproducibility. As all variables were direct indicators rather than latent constructs, no separate measurement model was involved in PLS-SEM.

2.3.4. Data Discretization Methods for GeoDetector

The NDVI, meteorological variables, LCCs, Aspect, Slope, elevation, soil properties and surface SM were incorporated into the GeoDetector model to quantify the spatial explanatory power of various environmental factors on the surface SM distribution. GeoDetector emphasizes the hierarchical structure and heterogeneity of spatial attributes. A critical step before conducting GeoDetector is converting continuous data into discrete forms. During this transformation, the number of classifications is crucial, since excessive classifications can increase computational complexity and introduce redundancy, whereas too few classifications may overlook spatial diversity. Therefore, selecting an appropriate discretization method and determining the optimal number of classifications are essential in GeoDetector analysis [75]. The discretization can be implemented based on expert knowledge or automatically generated based on the characteristics of the data distribution. The main discretization methods include the Equal Interval (EI) method, Quantification (QU) method, Natural Breaks (NB) method, Geometric Interval (GI) method, and Standard Deviation (SD) method [35,76]. By applying multiple discretization methods, we reduce the potential bias associated with any single scheme and better capture the spatial heterogeneity of each variable, as suggested by Song et al. (2020), which also enables a preliminary assessment of the sensitivity of the GeoDetector results to different discretization choices, mitigating the risk of overestimating the explanatory power due to a specific classification scheme [76].

In this study, the above data discretization methods were used to discretize 8 continuous environmental variables: TE, Elev., NDVI, PRE, RH, SSR, T, and WS. By comparing various discretization methods, the optimal method for data discretization was selected based on the number of cut points and the discretization technique that yielded the maximum q-value. The Slope, referring to the Second National Land Survey Technical Regulations, was divided into Flat (less than 2°), Gentle (2–6°), Moderate (6–15°), Steep (15–25°), and Very Steep (greater than 25°). Aspect was conventionally divided into 9 categories: Flat, North, Northeast, East, Southeast, South, Southwest, West, and Northwest [77]. The BULK and SOC indicators each contain many zero values and a relatively small number of unique categories (178) in the study area, which makes typical automatic discretization methods unsuitable. Following Hu et al. (2025), 0 values were treated as a separate class, and the remaining data were discretized into seven intervals using the EI method [78]. The 0 values, representing non-soil areas, were excluded when analyzing their effect on SM. This approach preserves adequate representation of non-zero values and allows a practical sensitivity assessment under limited-category scenarios.

2.3.5. Geographical Detector

GeoDetector offers four core functionalities, i.e., the factor detector, interaction detector, ecological detector, and risk detector [79]. The factor detector helps identify whether a particular factor is associated with the spatial distribution of a phenomenon. The interaction detector evaluates whether the combination of factors amplifies or diminishes their influence on this distribution. The ecological detector assesses whether the impact of interacting factors is statistically significant, typically expressed by the F-value. Lastly, the risk detector quantifies the risks tied to specific factors, generally measured by the T-value [32].

In this study, the factor detector was used to identify the dominant variables of surface SM spatial heterogeneity. The dominant variable is determined by the explanatory power (q-value) of a variable (X) to the spatial heterogeneity of surface SM (Y). The equation for q is as follows:

$$q=1-{\displaystyle \frac{{\displaystyle {\sum }_{j=1}^{\mathrm{i}}{N}_j{\sigma }_j^2}}{N{\sigma }^2}},$$

(3)

where j = 1, 2, …, i represents the stratification of variable Y or X; N_j and N denote the number of units in layer j and in the entire study area, respectively; ${\sigma }_j^2$ and σ² correspond to the variance of the Y value within layer j and across the whole area. The q-statistic ranges from 0 to 1, where a larger q-value indicates a stronger explanatory power of variable X on Y, and a smaller q-value suggests a weaker relationship.

The interaction detector was employed to assess whether the explanatory powers of two variables were enhanced, weakened, or independent compared to that of a single variable. The interaction between two variables was classified into five categories, namely, Weaken_nonlinear, Weaken_univariate, Enhance_bivariate, Independent, and Enhance_nonlinear, according to the q-values of individual and interactive variables for SM [79]. Specifically, the q-values for two variables, X₁ and X₂, with respect to Y were first calculated (q(X₁) and q(X₂)). Next, the q-value of the interaction—formed by the intersection of overlaying variables X₁ and X₂—was calculated (q(X₁ ∩ X₂)). Finally, by comparing q(X₁ ∩ X₂) with q(X₁) and q(X₂), the interaction between two factors can be classified into five categories, as shown in

Table 2. Categories of interactions between two driving variables and the response variable.

Judgment Criteria	Interaction Type
q(X1 ∩ X2) = q(X1) + q(X2)	Independent
q(X1 ∩ X2) > q(X1) + q(X2)	Enhance_nonlinear
q(X1 ∩ X2) < Min(q(X1), q(X2))	Weaken_nonlinear
q(X1 ∩ X2) > Max(q(X1), q(X2))	Enhance_bivariate
Min(q(X1), q(X2)) < q(X1 ∩ X2) < Max(q(X1), q(X2))	Weaken_univariate

.

The risk detector calculates the mean surface SM for different intervals of environmental factors in this study and determines whether significant differences exist in the mean surface SM between two intervals. These differences are assessed using the t-statistic.

$$t_{{\overline{Y}}_{h=1}-{\overline{Y}}_{h=2}}={\displaystyle \frac{{\overline{Y}}_{h=1}-{\overline{Y}}_{h=2}}{{\left[\frac{Var({\overline{Y}}_{h=1})}{n_{h=1}}+\frac{Var({\overline{Y}}_{h=2})}{n_{h=2}}\right]}^{1/2}}},$$

(4)

where ${\overline{Y}}_h$ represents the mean of the surface SM within interval h, n_h is the sample size in interval h, and Var stands for variance [32]. If the difference in the mean surface SM between two intervals passes the t-test, it is considered that the surface SM shows significantly different sensitivities between these two intervals. When the mean surface SM reaches the maximum value within a specific interval of an environmental factor, surface SM exhibits high sensitivity within that interval. Conversely, when the mean surface SM reaches the minimum value within a specific interval, surface SM exhibits low sensitivity within this interval. In the remaining intervals, the surface SM exhibits moderate sensitivity. GeoDetector in this study was conducted using the ‘Geographical Detector’ plugin in QGIS (v3.34.3), and the statistical analysis was performed using Python (v3.11).

2.3.6. Integrative Importance of Environmental Drivers of the Spatiotemporal Variability of Surface SM

The dominant drivers of the surface SM variability may differ between temporal and spatial dimensions, necessitating the establishment of a weighted hierarchy of meteorological variables and the NDVI based on their respective temporal and spatial contributions. To construct an integrated index that comprehensively interprets surface SM dynamics, we first normalized the total effects from the PLS-SEM and q-value of variable X from GeoDetector to ensure comparability, as recommended for composite indicator construction [80].

Specifically, each variable’s q-value and total effect was normalized by its method-specific sum to obtain comparable 0–1 values, and the normalized temporal and spatial contributions were then summed to produce the unified spatiotemporal weight index (W_ts). This additive approach follows standard decision-level fusion principles [81], integrating complementary temporal and spatial information while preserving each driver’s relative importance. The mathematical formulation of the W_ts index is given as follows:

$$\begin{array}{l}{W}_{time}={\displaystyle \frac{\left|E_{total}\right|}{SUM(\left|E_{total}\right|)}}\\ W_{space}={\displaystyle \frac{q}{SUM(q)}}\\ W_{ts}=W_{time}+W_{space},\end{array}$$

(5)

where W_time represents the weight of environmental factor W on the temporal variation of the surface SM; E_total is the total effects of W on the temporal variation of the surface SM obtained by PLS-SEM; SUM(E_total) stands for the sum of the total effects for all environmental factors; W_space donates the weight of environmental factor W on the spatial variation of the surface SM; q is the q-value of W on the spatial variation of the surface SM obtained by factor detector; and W_ts represents the spatiotemporal weight of W.

After normalization and aggregation, the temporal and spatial contributions captured are combined into a single metric W_ts, synthesizing the spatial heterogeneity and temporal causal strength. Consequently, the unified index furnishes a comprehensive quantification of each driver’s relative importance to the spatiotemporal dynamics of surface SM, providing a physically interpretable metric that reflects the influence of environmental factors on both the spatial heterogeneity and temporal variability of SM.

3. Results

3.1. Spatiotemporal Dynamics of Surface SM and the Environmental Factors

The spatial distribution of the multi-year average surface SM and the environmental factors (NDVI, PRE, SSR, T, WS, RH and TE) in the TS is shown in Figure 3. The mean value of the surface SM was 0.21 m³·m⁻³, with greater values observed in the central part. The NDVI, PRE, and RH exhibited similar spatial patterns, with mean values of 0.16, 218 mm and 53%, respectively. SSR, T, and TE were high in the eastern and southern parts, with mean values of 141 W·m⁻², 280 K and −309 mm, respectively. WS was high in the eastern part, with mean values of 4.21 m·s⁻¹. The spatial coefficients of variation (CVs) for the multi-year average surface SM, NDVI, PRE, SSR, T, WS, RH, and TE were 33%, 69%, 55%, 14%, 2%, 24%, 22%, and 48%, respectively. It indicated that all variables exhibited moderate spatial variability except for T.

Table 3. Multi-year average value for surface soil moisture and environmental factors across different dimensions.

	Surface SM (m3·m−3)	NDVI	PRE (mm)	SSR (W·m−2)	T (K)	WS (m·s−1)	RH (%)	TE (mm)
The entire area	0.21	0.16	218	141	280	4.21	53	−309
Cropland	0.24	0.27	247	137	280	3.87	56	−372
Forestland	0.26	0.33	309	128	277	3.78	62	−437
Grassland	0.24	0.21	281	133	277	3.88	59	−383
Bare land	0.17	0.10	153	151	283	4.59	47	−235
Low elevation	0.16	0.14	143	149	286	4.82	46	−223
Medium elevation	0.21	0.18	217	143	281	4.16	53	−323
High elevation	0.27	0.09	343	118	269	3.59	66	−346

shows the mean values of surface SM and environmental factors across different LCCs and elevation gradients. Among all LCCs, forestland recorded the highest surface SM (0.26 m³·m⁻³), NDVI (0.33), and RH (62%), along with the lowest T (277 K) and SSR (128 W·m⁻²). In contrast, bare land showed the lowest surface SM (0.17 m³·m⁻³), NDVI (0.10), and RH (47%), accompanied by the highest SSR (151 W·m⁻²) and relatively high temperature (283 K). Each variable also varied markedly with elevation. Surface SM generally increased with elevation. PRE and RH showed a continuous increasing trend, rising from 143 mm and 46% at low elevation to 343 mm and 66% at high elevation, respectively. In contrast, T and SSR decreased with elevation. Both WS and TE also declined with increasing elevation. The NDVI increased slightly from low to medium elevation but dropped sharply at high elevation. Overall, LCC and elevation strongly affect surface SM and other environmental conditions.

From 2000 to 2022, the variations in annual surface SM and its environmental factors showed obvious spatial heterogeneity (Figure 4). In most areas (94%), surface SM exhibited an increasing trend. Further, there was a significant (p < 0.05) increase in 50% of the region, while only 0.62% exhibited a significant decrease. The NDVI showed an increasing trend in most areas (79%), and 38% of the study area experienced a significant rise in the NDVI, primarily concentrated in the eastern and southern parts. Only 2% of the region showed a significant decrease in the NDVI. PRE exhibited a decreasing trend in most areas (78%), while only 0.6% of the study area showed significant changes. Although SSR showed an increasing trend in most areas (90%), only 2% experienced a significant increase. T exhibited an increasing trend in 85% of the study area, but it was only significant in 20% of the region, which was primarily concentrated in the northeastern and central parts. WS showed an insignificant decreasing trend across 72% of the study area, with significant increases or decreases observed in only 1% of the region. RH exhibited a decreasing trend in 99% of the area, with a significant decrease in 50% of the region, primarily concentrated in the central and western parts. TE exhibited an increasing trend in 61% of the region, and the area ratio was 11% for a significant increase, primarily concentrated in the central and northern parts. Meanwhile, 38% of the study area showed a decreasing trend in TE, and 4% exhibited a significant decrease, which was mainly concentrated in the central part. In summary, the surface SM and NDVI showed predominantly increasing trends, while RH generally declined. By contrast, PRE, T, SSR, TE, and WS exhibited weaker or less significant changes, reflecting varied spatial patterns of trends among the examined variables.

Table 4. Multi-year variation trend (per decade) for surface soil moisture and environmental factors across different dimensions.

	Surface SM (m3·m−3)	NDVI	PRE (mm)	SSR (W·m−2)	T (K)	WS (m·s−1)	RH (%)	TE (mm)
The entire area	0.011 **	0.007 **	−6.121	1.536	0.326 *	−0.017	−1.483 **	6.980
Cropland	0.012 **	0.014 **	−8.541	1.859	0.313	−0.009	−1.480 **	13.296 *
Forest	0.009 **	0.006	−10.756	2.254	0.506 **	−0.011	−1.358 **	−3.270
Grassland	0.010 **	0.005 *	−8.407	1.943	0.360 *	−0.017	−1.411 **	0.874
Bare land	0.012 **	0.007 **	−3.440	1.097	0.302	−0.020	−1.560 *	10.989
Low elevation	0.012 **	0.011 **	−7.046	1.236	0.401 *	−0.008	−1.585 **	22.771 *
Medium elevation	0.012 **	0.007 **	−6.154	1.505	0.309	−0.017	−1.546 *	5.733
High elevation	0.004 *	0.005 **	−2.960	2.209	0.331	−0.035	−0.939 *	−8.267

*, ** indicate significance at the p < 0.05 and p < 0.01 levels, respectively.

shows the variation trends of the annual surface SM and environmental factors across different LCCs and elevation gradients. Across all LCCs, the surface SM and NDVI showed consistently significant increasing trends (p < 0.01), whereas RH generally exhibited significant negative trends (p < 0.01). PRE and WS showed slight declining trends, though not statistically significant, while SSR showed a slight, non-significant increasing trend. T showed an overall upward trend, which was significant only in forest and grassland areas, and TE increased significantly only in cropland. Across low to high elevations, the surface SM and NDVI consistently showed significant increasing trends (p < 0.05), while RH displayed significant decreases (p < 0.05). PRE, SSR, and WS showed no significant changes. T and TE increased significantly (p < 0.05) only at low elevation. For the entire study area, surface SM, NDVI and T showed significant increases (p < 0.05), while the increase in SSR and TE was not significant. Moreover, RH decreased significantly (p < 0.01), but the decrease in PRE and WS was not significant. These results indicate that the surface SM and NDVI generally increased across LCCs and elevations, RH consistently declined, T showed moderate increases, and PRE, SSR, WS, and TE exhibited mostly non-significant changes.

3.2. The Impacts of Meteorological Factors and Vegetation Dynamics on the Temporal Variations in Surface SM

PLS-SEM demonstrated acceptable performance (p < 0.05, R² > 0.33, NFI > 0.90 and SRMR < 0.08) in quantifying the direct, indirect, and total effects of meteorological and vegetation factors on the temporal variations in surface SM across multiple dimensions (

Table 5. Assessment of PLS-SEM model using SRMR and NFI at different dimensions.

Model Fit indices	The Entire Area	Low	Medium	High	Cropland	Forest	Grassland	Bare Land
SRMR	0.05	0.08	0.03	0.08	0.08	0.07	0.04	0.04
NFI	0.95	0.92	0.92	0.93	0.90	0.92	0.95	0.95

PLS-SEM—Partial Least Squares Structural Equation Modeling, SRMR—Standardized Root Mean Square Residual, and NFI—Normed Fit Index.

). The direct, indirect, and total effects of these factors on surface SM variations exhibited significant differences (Figure 5). For the entire study area, the NDVI was the strongest direct positive driver of surface SM, whereas T represented the primary direct negative driver. Meanwhile, PRE and RH influenced surface SM primarily through the NDVI-mediated pathway (Figure 5a). In addition, T exerted an overall positive indirect effect on surface SM through two pathways, by enhancing the NDVI and suppressing RH.

The effects of climate and vegetation greenness on the surface SM variability differed along the elevation gradient (Figure 5 and Figure 6). Across low, medium, and high elevations, the NDVI was consistently the strongest direct and total driver of surface SM variability. At low and medium elevation zones, the primary pathway of PRE operated indirectly through enhancing the NDVI, whereas at high-elevation zones, PRE acted as a direct positive driver of surface SM. Across all elevation zones, T exerted an important indirect influence via the NDVI-mediated pathway, although it served as a direct negative driver in medium and high elevations. Additionally, SSR exerted a direct negative effect on surface SM only at high elevation, while at low and medium elevations it influenced surface SM indirectly through T.

Across different LCCs, the NDVI consistently exerted the strongest total effect on surface SM dynamics, primarily through its direct positive effect. In cropland, T acted as an important but dual-pathway driver, indirectly enhancing surface SM by promoting the NDVI while simultaneously exerting a direct negative effect. WS emerged as a distinct direct positive driver. In forestland, the total effects of T, SSR and NDVI on surface SM were all significant. The primary pathway of T operated indirectly through NDVI enhancement. Although SSR exerted a stronger direct negative effect than the direct positive effect of the NDVI, its overall total effect was weaker than that of the NDVI, because SSR also indirectly promoted surface SM via T and the NDVI. In grassland, the driver hierarchy was similar to that in forestland, with the NDVI as the dominant driver, while PRE contributed mainly through the NDVI-mediated pathway. In bare land, T was the primary direct negative driver, although it also exerted an indirect positive influence via the NDVI, similar to the indirect effects of RH and PRE.

3.3. Individual and Interactive Effects of Environmental Factors on Surface SM Based on the GeoDetector

In the previous section, PLS-SEM was employed to quantify the direct, indirect, and total effects of environmental factors on the interannual variability of surface SM from a temporal perspective, thereby identifying the dominant drivers and primary pathways. Although this temporal analysis elucidates the dynamic regulatory mechanisms, it does not explicitly reveal how these drivers contribute to the spatial heterogeneity. To complement the temporal findings, this section applies the GeoDetector approach to evaluate the spatial explanatory power of environmental factors and their interactions on shaping the spatial patterns of surface SM.

3.3.1. Optimal Data Discretization Method and the Corresponding Interval

Figure 7 and

Table 6. Optimized discretization methods and dispersion interval number for the multi-year average spatial continuous factors.

Variable	Elev. (m)	PRE (mm)	RH (%)	SSR (W·m−2)	T (K)	WS (m·s−1)	TE (mm)	NDVI
Discretization Method	QU	QU	GI	QU	QU	GI	NB	NB
Discrete Interval Number	The Starting Value of Each Interval
1	285	22	29	46	247	2.61	−709	0.01
2	847	82	33	115	271	2.96	−534	0.04
3	1255	108	37	126	276	3.35	−450	0.08
4	1552	146	41	134	279	3.79	−383	0.12
5	1849	179	45	141	282	4.29	−320	0.17
6	2288	221	51	148	284	4.85	−248	0.22
7	2842	267	56	153	286	5.49	−176	0.28
8	3436	316	63	158	287	6.22	−108	0.34
9		382	70	165				0.41

QU—quantification method, GI—geometric interval method, and NB—natural breaks method.

demonstrate the discretization process and the optimal results of eight continuous multi-year average variables (i.e., Elev., PRE, RH, SSR, T, WS, TE, and NDVI). The results indicated that the optimal number of intervals for Elev. and T was 8, with the discretization method chosen as QU. For TE, NDVI, PRE, RH, SSR, and WS, the optimal number of intervals was 8, 9, 9, 9, 9 and 9, respectively. The corresponding discretization methods were NB, NB, QU, GI, QU and GI, respectively. Similarly, we discretized all annual continuous environmental factors from 2000 to 2022 and selected the discretized data with the optimal interval as the input for GeoDetector to analyze the driving factors of SM spatial heterogeneity.

Table 7. Optimized methods and interval numbers for the discretization of spatial continuous environmental factors from 2000 to 2022.

Year	Elev.	PRE	RH	SSR	T	WS	TE	NDVI
2000	NB6	QU6	NB7	QU5	QU5	NB6	NB8	NB9
2001	QU8	QU9	NB7	QU6	QU7	GI9	QU8	NB9
2002	QU8	QU7	NB8	QU8	NB7	NB7	NB8	NB9
2003	QU7	QU7	EI9	QU7	QU6	NB6	NB7	NB8
2004	NB6	QU7	QU8	QU8	QU8	NB8	NB7	NB9
2005	QU7	QU7	NB7	QU7	QU7	NB7	NB9	NB9
2006	NB6	QU7	GI8	QU7	NB5	NB6	NB8	QU5
2007	NB6	NB7	NB7	QU7	QU6	NB6	QU7	NB9
2008	NB6	QU7	NB7	QU7	QU6	NB6	QU7	NB9
2009	NB6	QU8	GI7	QU7	QU6	QU6	QU7	NB8
2010	NB6	QU7	QU7	QU7	QU6	NB7	QU7	QU6
2011	QU7	QU6	NB7	QU7	QU7	NB7	QU7	NB9
2012	NB6	QU6	NB5	QU7	QU5	NB7	QU8	NB9
2013	NB6	QU5	QU5	QU6	QU5	QU5	NB8	NB9
2014	QU7	QU7	NB7	QU6	QU5	NB7	NB8	NB9
2015	QU5	QU7	GI8	QU7	QU7	NB7	NB5	NB9
2016	QU7	QU7	NB8	QU7	QU6	GI8	EI7	NB9
2017	QU8	QU7	NB5	QU7	QU8	NB8	QU7	NB9
2018	NB6	QU8	NB8	QU6	QU5	NB7	NB8	NB8
2019	NB6	QU8	NB8	QU7	NB5	NB6	QU8	NB9
2020	NB6	QU6	GI8	QU7	QU8	NB7	NB6	NB9
2021	QU7	QU8	NB7	QU6	QU7	NB7	QU8	NB9
2022	QU8	QU6	NB8	QU7	QU7	NB7	QU8	NB9

shows the optimal methods and interval numbers for discretizing each continuous variable from 2000 to 2022. The results indicated that the optimal discretization methods and intervals varied across years. QU was the most frequently used method (98 times), followed by NB (84 times), while GI and EI were less used (8 and 2 times, respectively), suggesting that QU and NB were more suitable for discretizing spatial continuous environmental factors in the TS. The most common optimal interval number was 7, followed by 8 and 6, indicating that more intervals did not always improve the discretization results.

3.3.2. Impact of Individual Factors on the Spatial Heterogeneity of Surface SM

The factor detector results from 2000 to 2022 provided the contributions (q-values) of various driving factors to the spatial heterogeneity of surface SM (Figure 8), with all q-values passing the significance test (p < 0.01). The mean q-values of individual factors during the study period were ranked as follows: RH (0.72) > PRE (0.71) > SSR (0.65) > WS (0.55) > T (0.49) > TE (0.49) > SOC (0.43) > Elev. (0.33) > NDVI (0.32) > LCC (0.26) > BULK (0.19) > Slope (0.19) > Aspect (0.01), which indicated that RH and PRE showed the highest contributions to the spatial heterogeneity of surface SM. SSR and WS also played significant roles. The q-values of T, TE and SOC were smaller than those of the previous four factors, yet they still played essential roles. The influence of Elev., NDVI and LCC was less pronounced compared to the meteorological factors mentioned above. The effects of BULK and Slope on SM were also relatively small during the study period. Although the q-value of Aspect was much lower than that of other factors, its t-test p-value remained below the 0.01 threshold, indicating that although Aspect had a relatively minor impact on the spatial heterogeneity of surface SM, it still exerted a statistically significant influence. The individual influence of driving factors on the multi-year average surface SM was similar to the results above.

To assess the temporal variations in the spatial explanatory power of different drivers, we calculated the change rate of the q-values for all drivers from 2000 to 2022 (Figure 8). The results indicated that the contribution of TE showed an increasing trend; that of Aspect, NDVI, and LCC exhibited weak change; and the contribution of other driving factors showed a decreasing trend. Among these drivers, SSR, RH and Elev. showed the greatest change rate of contribution (with a trend for the q-values of −0.003 year⁻¹). Despite the temporal variability in contribution, PRE, RH, and SSR consistently emerged as the dominant factors shaping the spatial distribution of surface SM.

3.3.3. Interaction Between Pairwise Factors on the Spatial Heterogeneity of Surface SM

The q-values of interactions (q(X₁ ∩ X₂)) for pairwise factors in Figure 9 demonstrated that the interaction between any two factors enhanced their individual influence on surface SM, suggesting that the spatial heterogeneity of surface SM was jointly controlled by various drivers rather than a single driver.

During the study period, the interaction between RH and WS was the strongest throughout the entire period, with q-values ranging from 0.74 to 0.85 and an increase of 7–20% compared with the RH factor alone; the interaction between PRE and WS was also substantial, with q-values ranging from 0.69 to 0.82 and an increase of 8–26% compared with PRE alone; and the interaction between RH and PRE was similarly pronounced, with q-values ranging from 0.67 to 0.81 and an increase of 3–9% compared with RH alone. These three interactions represent the dominant interactions across the study period. Each year, the interactions between RH (or PRE) and any other driving factor remained consistently high, with q-values exceeding 0.65 for RH-related interactions and 0.63 for those involving PRE. The interaction between SSR and any other factor was also great, with q-values in all years exceeding 0.54, and the interaction with RH was the largest. Aspect mainly showed Enhance_nonlinear interactions with LCC, BULK, NDVI, Elev., and T, especially for LCC and BULK, which were consistently observed each year. The q-values of interactions for the multi-year average surface SM (Figure 9x) were consistent with those for individual years. Most interactions belonged to Enhance_bi-interaction, and the interactions with RH or PRE were stronger. Meanwhile, Aspect exhibited Enhance_nonlinear interactions with some factors.

4. Discussions

4.1. Spatial Sensitivity of Surface SM to Driving Factors

The strong spatial explanatory power of RH, PRE and SSR across both individual and interactive analyses underscored their critical roles in regulating the surface SM spatial distribution in the TS. On the other hand, the results of the factor detector and interaction detector indicated that GeoDetector analysis on multi-year average data yielded similar results to those of each year. Therefore, to determine the sensitivity of the surface SM to its main driving factors at the spatial scale, the risk detector results of the multi-year average data were used, and then the average surface SM values for different intervals of RH, PRE and SSR were compared (Figure 10). The results indicated that for PRE and RH, pixels with low values of driving variables showed low surface SM values, and vice versa. Surface SM was highly sensitive to PRE where PRE was greater than 316 mm (Table 6). This phenomenon could be partly attributed to the fact that low precipitation induced soil particle aggregation, leading to surface crust formation and hindering water infiltration, thus limiting an SM increase [82,83]. However, pixels with low values of SSR showed high SM values. To further identify the sensitive areas of surface SM to RH, PRE and SSR, the spatial distribution of the surface SM with different sensitivities is also shown in Figure 10. As a result, RH, PRE and SSR all showed low sensitivity in the northeast, south, and southwest in the TS and high sensitivity in the central high alpine area. The sensitivity of SM to RH and SSR suggests that low soil moisture was present with high SSR and low RH. In the future, global warming will continue [84], which may lead to high SSR and low RH in the study area, resulting in larger SM deficits.

4.2. Impacts of Environmental Factors on Surface SM Spatiotemporal Variations

To explore the mechanisms underlying the SM variability, both the spatial heterogeneity and temporal responses of environmental factors were examined. The spatial CV of environmental factors varied substantially, yet their explanatory power for SM spatial heterogeneity (q-values) did not show a direct correspondence with CV magnitude. This suggests that large spatial variability alone does not necessarily imply strong control over the SM spatial pattern. Instead, factors closely linked to soil water input and hydro-climatic processes exhibited the highest contributions. In particular, PRE (q-values = 0.71) and RH (q-values = 0.72) showed strong explanatory power, indicating that water supply conditions play dominant roles in shaping the spatial differentiation of surface SM. SSR (q-values = 0.65) and WS (q-values = 0.55) also contributed considerably, likely through their regulation of evapotranspiration processes. These findings are consistent with previous studies demonstrating that precipitation governs SM variability [7] and that RH regulates SM spatial patterns [85]. The spatial CV of TE across the study area was 48%, placing it at an upper-moderate level among all environmental factors, whereas trend analysis indicated that only 11% of the area exhibited a significant temporal change in TE during the study period. This explains why TE exerts a relatively strong influence on the spatial distribution of SM (q-value = 0.49), but its temporal effect on SM is limited. Overall, the spatial pattern of SM appears to be primarily controlled by hydro-climatic drivers rather than by the spatial differences in environmental factors.

Numerous studies have shown that precipitation and temperature are two important factors influencing SM [18,69,86]. However, vegetation greenness and precipitation were identified as two dominant positive drivers of annual surface SM dynamics in our study, and T only exhibited a weak negative effect since the direct negative effect of T was counteracted by its indirect positive effect. More, PRE indirectly influenced surface SM by promoting the NDVI in this region, except for high elevation, cropland and forestland.

The impact of vegetation on SM is complex. Some studies showed that increased vegetation greenness would lead to soil drying [87,88], while other studies pointed out that vegetation recovery could enhance surface SM [89]. Our results were in line with the latter. It can be explained by the fact that vegetation cover reduces soil evaporation by shading the soil surface, which lowers the soil temperature, enhances local humidity, and stabilizes SM through bio-geophysical and microclimate feedback, particularly in semi-arid regions [90,91]. Moreover, vegetation growth can enhance soil structure and infiltration, allowing more water to be retained in the soil [92,93]. Additionally, vegetation promotes ecohydrological feedback by altering the microclimate, such as enhancing local humidity and modifying wind patterns, further enhancing SM retention. This feedback mechanism is particularly evident in regions with limited water availability, as vegetation acts as both a moderator of evapotranspiration and a stabilizer of SM [94,95]. Actually, vegetation can regulate SM by suppressing evaporation [94].

The SM–PRE relationship is nonlinear and influenced by rainfall characteristics, such as event size and intensity, rather than total precipitation alone. Multiple small rainfall events are often less effective in replenishing SM because much of the water rapidly evaporates, whereas fewer large events substantially enhance infiltration and soil water storage [7,96]. Considering that the extreme precipitation events occur more often under climate warming [84], SM can increase even if total precipitation does not show significant variations, as shown in this study.

In our study, T showed a negative direct effect on SM but a positive indirect effect via vegetation greenness. It can be explained by the fact that T can promote vegetation growth, which then improves soil structure, infiltration, and SM [90,91,92,93,97]. Under ongoing warming, the indirect effect of T via vegetation may partially stabilize SM in vegetated regions, enhancing ecosystem resilience. However, in sparsely vegetated or high-elevation areas where vegetation buffering is weak, warming may lead to greater vulnerability to drying, thereby amplifying drying–greening contrasts.

RH was also an important moisture-related variable impacting the spatial distribution of surface SM. Actually, RH showed important individual impacts on the spatial distribution of surface SM. The higher the RH, the more sensitive the surface SM was to it. The reason is that when RH is above 60% and the air temperature–dew point difference is below 8 °C, conditions are more favorable for the formation of soil condensation water, which increases the surface SM value [98]. However, in terms of temporal variation, RH only exhibited significant positive effects on SM in bare land. This may be attributed to the vapor adsorption of soil from atmospheric moisture in bare land [98,99].

In addition to moisture-related factors, WS also plays an important role in regulating the surface SM spatial pattern through its influence on land–atmosphere water exchange. Moderate wind speeds can facilitate atmospheric moisture transport and condensation, thereby contributing to SM replenishment, whereas higher wind speeds enhance soil evaporation and plant transpiration, resulting in increased SM loss [11]. However, as WS showed no significant temporal variation during the study period in the TS, its influence on the annual SM variation was limited except for cropland. Actually, the SM dynamics are more sensitive to wind in agricultural land [100,101].

SSR influences surface SM through its control on the surface energy balance [102]. Spatially, stronger radiation corresponds to a lower SM. Temporally, the effect of SSR varies with elevation and vegetation: at high elevations, SSR tends to reduce SM directly, while in densely vegetated areas such as grassland and forest, its influence can be partially offset via temperature-induced vegetation growth [103], which promotes SM retention.

Soil properties also played a crucial role in the spatial distribution of SM in the TS, given that the individual impact of SOC and BULK on SM exceeded that of Elev., Aspect, and Slope, and the individual impact of SOC was also greater than that of LCC. These findings differed from those in the Heihe River Basin, where LCC and elevation were identified as the primary driving factors under low SM conditions, while the NDVI and T became the dominant factors when SM was relatively high [69].

4.3. Integrating Temporal and Spatial Explanatory Power of Environmental Drivers on Surface SM

As shown in Figure 11, for the entire TS region, the integrative impacts of the NDVI, PRE and RH on the spatiotemporal variations of annual surface SM were greater than those of other factors. For the NDVI and PRE, their impacts on the temporal variations were greater than those on the spatial variations. In contrast, RH exhibited a stronger spatial explanatory power. Similarly, the spatial heterogeneity of T also exerted a substantial influence on the spatial pattern of SM. However, due to its weaker temporal explanatory power, the integrative impact of T on the spatiotemporal variation in surface SM was smaller than that of RH. For TE, SSR and WS, although they exerted substantial control over the spatial distribution of surface SM, their overall influence on the spatiotemporal distribution of SM was minor due to their negligible temporal impacts. It was probably due to the fact that surface SM was mainly affected by TE at an annual time scale [96], and strong interactions among TE, SSR, and WS were evident, as shown in Figure 9. These findings suggested that despite numerous studies highlighting the crucial role of meteorological factors and vegetation in the spatiotemporal dynamics of surface SM [104], there were disparities in the magnitude and mode among different factors.

By quantifying the relative contribution of each driver through the W_ts index, which integrates the impact on spatial heterogeneity and temporal dynamics, we were able to identify the key climatic and vegetation factors controlling SM variability. We also developed a scheme of relative interpretability to reveal the spatiotemporal driving mechanisms, as demonstrated in integrated GeoDetector-regression frameworks [105]. However, the spatial and temporal contributions were assumed to be equally weighted and combined additively, without accounting for potential nonlinear interaction or physical processes.

4.4. Limitations and Prospects

This study presents important findings on the spatiotemporal heterogeneity of surface SM in the TS, as well as the corresponding impact pathways of climate, vegetation, topography and soil factors. However, there were still some limitations in this study. First, SM is a complex variable influenced by both natural and human factors, such as grazing intensity and agricultural irrigation [106,107,108]. However, anthropogenic regulation was not explicitly quantified in this study, which may introduce uncertainties in distinguishing the relative contributions of climatic and environmental controls at local scales. Second, there are great uncertainties in the soil moisture dataset in the mountainous areas [9]. Data uncertainties related to terrain may affect the results. Third, the TS receives significant snowfall during the winter. As a form of solid water resource, snow may not have an immediate effect on SM, but rather, exhibits a delayed impact [109,110]. Since this study primarily used annual mean or cumulative values, it is limited in its ability to capture the effect of seasonal snow and other short-term processes on SM. Finally, the assumption of time-invariant soil and terrain properties may limit the ability to capture the impact of land degradation or erosion over the study period. Therefore, future research should investigate the influence of various human factors, as well as snow accumulation and soil property changes, on the SM variability in the TS based on SM datasets with greater accuracy.

5. Conclusions

This study examined the spatiotemporal dynamics of surface SM driven by climate, vegetation, soil and topography factors in the Tianshan Mountains. It revealed how these environmental factors influenced the spatiotemporal heterogeneity of surface soil moisture, by applying PLS-SEM and GeoDetector. The main findings were as follows:

(1)

The multi-year average surface soil moisture in the Tianshan Mountains exhibited moderate spatial variability, with great values in the central high-elevation area dominated by forest and grassland and low values in the low-elevation area dominated by bare land. From 2000 to 2022, surface SM and NDVI showed significant increasing trends across different LCCs and elevation gradients, RH exhibited a significant decreasing trend, and other environmental factors showed weak changes.

(2)

NDVI exerted the most pronounced positive direct effect on surface SM dynamics, while PRE mainly exerted an indirect positive effect by promoting vegetation greenness. In contrast, T showed a negative direct effect on surface SM, which was offset by its positive indirect effect via vegetation greenness, highlighting the importance of vegetation–climate feedbacks in regulating SM dynamics. SSR influenced annual surface SM dynamics mainly via T, while RH influenced it via PRE. Notably, WS exerted a significant direct effect on surface SM only in cropland.

(3)

The individual effects of RH, PRE, SSR, WS, TE and T on the spatial variability of surface SM were greater than those of other factors, particularly RH, PRE, SSR. The interactive effects of these three factors with other factors were also strong, particularly the interaction between RH/PRE and WS, implying that WS regulates the surface SM spatial pattern via its influence on land–atmosphere water exchange. The spatial sensitivity of SM to RH and PRE increased with their values, but 316 mm was the upper threshold in terms of SM response to annual PRE. SM was highly sensitive to both RH and PRE in the central TS.

(4)

NDVI and PRE were the primary environmental drivers of surface SM spatiotemporal dynamics, followed by RH, SSR, T, TE, and WS. NDVI and PRE mainly impacted the temporal dynamics of surface SM, while SSR, T, TE and WS primarily affected its spatial patterns. These findings suggest that soil moisture can be enhanced through ecological and water resources management.

This study provides valuable insights into the impact pathways and underlying mechanism of environmental factors on the spatiotemporal heterogeneity of surface soil moisture in mountainous regions, which offers new clues for mountain hydrology and agropastoral water resource management in similar regions around the world.