Exploring the Main Driving Factors for Terrestrial Water Storage in China Using Explainable Machine Learning

Abstract

Terrestrial water storage (TWS) is a critical component of the hydrological cycle and plays a key role in regional water resource management. The launch of the Gravity Recovery and Climate Experiment (GRACE) satellite mission in 2002 has provided precise measurements of TWS, enabling systematic investigations into its spatial pattern and driving mechanisms. However, a comprehensive evaluation of the spatial drivers of TWS variations across China is still lacking. In this study, we employed a robust machine learning model to capture the spatial patterns of TWS in China and further applied the Shapley Additive Explanations (SHAP) method to disentangle the individualized effects of hydroclimatic variables. Our findings reveal that precipitation is the dominant driver in northern and southern China, while soil moisture and snow water equivalent are key contributors on the Tibetan Plateau. In northwestern China, air pressure and groundwater runoff are the main influencing factors, whereas temperature shows a pronounced negative effect. Importantly, most variables demonstrate non-monotonic influences: in particular, we found that the importance of precipitation diminishes beyond a certain threshold, and surface pressure shifts sharply toward a negative impact. The explainable machine learning framework demonstrated strong adaptability in identifying complex drivers of TWS, offering a powerful methodological advancement for exploring TWS dynamics and providing valuable insights for water resource management in China.

1. Introduction

In response to the escalating challenges of water resource scarcity, the development of effective water resource management strategies has become a pressing research imperative. Within this context, terrestrial water storage (TWS)—an integrated measure encompassing surface water, soil moisture, groundwater, and snow/ice—has emerged as a critical research focus due to its fundamental role in water cycle dynamics [1]. The spatial heterogeneity of TWS dictates the distribution of water resource availability [2] and fundamentally impacts regional water security. Typically, although China is one of the countries with the richest water resources in the world in terms of total volume, the spatial distribution of water resources is highly uneven due to climatic and geographical factors. Many arid and semi-arid regions continue to suffer from water scarcity [3]. Therefore, an in-depth investigation into the spatial distribution of TWS in China and its climatic and hydrological driving mechanisms is of significant scientific and policy relevance for optimizing differentiated regional water allocation strategies and alleviating water scarcity issues in arid and semi-arid regions.

Since the launch of the Gravity Recovery and Climate Experiment (GRACE) satellites in 2002, highly accurate measurements of TWS have become available. This technological breakthrough has enabled extensive monitoring of TWS variations, often through integration with hydrological models, land surface models, and other analytical frameworks [4,5,6,7,8,9,10]. In particular, the feasibility of using GRACE satellite data to track long-term TWS changes in China has been validated in previous studies [11,12,13,14]. Additionally, to overcome the limited spatial resolution of GRACE data, subsequent research efforts have incorporated complementary approaches, such as land surface models and remote sensing techniques [15]. Among these, Global Land Data Assimilation System (GLDAS) 2.2 stands out by assimilating GRACE satellite observations to enhance both the spatial resolution and the accuracy of TWS estimates [16], thereby facilitating a more refined analysis of the driving factors and underlying mechanisms governing TWS variability. Currently, TWS has been observed to have declined globally at a rate of 1 cm per year over the past two decades [17], which is closely associated with the impacts of extreme weather events [18,19,20,21], with China experiencing a decline of about 2 mm annually. However, the primary factors driving TWS variations are highly region-specific. For example, Girotto et al. [6] identified groundwater extraction for irrigation as the primary cause of TWS depletion in India, while Kim et al. [22] discovered the influence of river storage variations on seasonal TWS changes in wet basins.

Similarly, despite these advances, substantial knowledge gaps persist, particularly in regions with complex topoclimatic conditions and pronounced spatial heterogeneity, such as those found across China [23,24]. Recent studies have begun to shed light on these complexities. For example, Li et al. [25] demonstrated that the dominant drivers of TWS in southern China are precipitation and runoff [26], while Xie et al. [27] reported that the TWS declines in northern China are largely attributable to anthropogenic activities, whereas increases in southern China are mainly driven by precipitation. Yang et al. [28] emphasized the significant roles of evapotranspiration (ET) and runoff in regulating TWS in southern regions. These findings point to two major limitations in current research. Firstly, most studies focus on a narrow set of climatic variables—typically no more than four, such as precipitation and evaporation—which may not comprehensively represent the full complexity of the hydrological cycle. Secondly, the predominant reliance on simple linear regression models limits the capacity to capture the spatial heterogeneity of TWS dynamics and fails to adequately address multicollinearity among input features, potentially leading to biased interpretations.

With the rapid advancement of machine learning (ML) technologies, there has been a growing trend toward leveraging ML’s superior capabilities in feature extraction and pattern recognition for hydrological studies, including applications to runoff, soil moisture, and groundwater prediction [29,30,31,32,33]. Artificial neural networks (ANNs) are well-suited for modeling nonlinear relationships and are resistant to noise, and their flexible structure makes them popular for tasks like flood forecasting and water quality modeling [34]. Support vector machines (SVMs) offer strong generalization and global optimal solutions, providing accurate and stable predictions even with small datasets, and they also train faster than ANNs and are often used for forecasting tasks, such as reservoir inflow prediction [35]. Ensemble models like random forests (RF) and boosting methods have shown strong performance and robustness in handling complex, high-dimensional data, while maintaining good interpretability. Deep learning (DL) excels at extracting features from complex and high-dimensional data and has been successfully applied to complex tasks like multisatellite data fusion, gap-filling, and prediction in data-scarce regions [36]. DL models can leverage advanced structures like convolutional layers and attention mechanisms to detect relevant or complementary information across multiple data sources, improve cross-validation, remove redundancy, and enhance prediction accuracy and reliability.

However, both traditional ML and recently developed DL approaches tend to function as “black-boxes”, making it difficult to reveal the underlying physical mechanisms of the models. This limitation restricts their application value in hydrological process diagnosis, causal inference, and scientific decision-making. To address this limitation, explainable ML (XML) techniques have emerged, aiming to enhance the transparency and interpretability of model outputs while preserving the strong predictive capabilities of ML models [37,38]. XML refers to a suite of methods designed to make the internal logic and output of complex models more understandable, enabling users to grasp the rationale be-hind predictions. This not only allows for more trustworthy and interpretable results, but also lays the groundwork for integrating data-driven modeling with hydrological domain knowledge. For this study, we selected an ML model, which has been shown to capture TWS patterns effectively without the added complexity and uncertainty of DL [39,40,41]. By pairing it with an XML framework, we achieved a synergistic balance between predictive accuracy and mechanistic insight, ultimately increasing the model’s value for both scientific research and practical water resource management.

Based on these motivations, the main contributions of this study are threefold: (1) the development of a machine learning model that effectively captures the spatial patterns of TWS using hydroclimatic variables; (2) the identification of dominant spatial drivers of TWS in China through XML methods; (3) the interpretation of the underlying mechanisms revealed by model explanations. Through this research, we aim to advance understanding of the hydrological system and provide scientific support for more effective and targeted water resource management strategies.

2. Materials and Methodology

2.1. Study Area

China, located in eastern Asia between 3.9°–53.5°N and 73.7°–135°E, spans nearly 9.6 million square kilometers and exhibits remarkable climatic and geographic diversity. Its extensive latitudinal range and complex topography give rise to distinct climate zones and a wide range of landscapes, leading to a pronounced variability in climate and TWS. For example, the difference between the subtropical monsoon climate in southern China and the temperate continental arid climate in the northwest establishes a marked southeast-to-northwest gradient in precipitation. Correspondingly, the landscape transitions from humid southeastern coastal plains, forests, and hills to increasingly arid mountains and deserts in the northwestern interior. In addition to natural climatic influences, intensive human activities—driven by a population exceeding 1.4 billion—have substantially reshaped the hydrological environment. Large-scale interventions, such as water conservancy projects and groundwater extraction, have profoundly altered the natural water cycle. Together, these natural and anthropogenic processes drive strong spatial heterogeneity in TWS, making China an ideal study area for investigating the drivers of TWS variations (Figure 1).

2.2. Data Source

The hydroclimatic data used in this study were derived from the GLDAS, developed by NASA’s Goddard Space Flight Center. GLDAS integrates satellite- and ground-based observations with advanced land surface models to generate high-resolution terrestrial hydroclimatic datasets with near-real-time capability and long-term consistency [14,42]. The Community Land Surface Model (CLSM), version 2.2, assimilates GRACE observations. This improves the consistency and accuracy of CLSM-derived TWS estimates, making them well-suited for long-term TWS analysis [43,44,45]. The CLSM TWS data are provided at a spatial resolution of 0.25° and a daily temporal resolution from 1 January 2003 to the present (1 September 2022, used in this study). Additionally, a multi-year mean was calculated from the TWS data using the Climate Data Operators software 2.2.0 to remove the temporal dimension and represent average spatial patterns. The original TWS data can be publicly accessed at https://daac.gsfc.nasa.gov/datasets/GLDAS_CLSM025_DA1_D_2.2/summary (accessed on 10 January 2023).

To capture the spatial distribution patterns of TWS, several key hydrometeorological variables were selected, including precipitation (P), surface air pressure (SP), air temperature (T), soil moisture (SM), groundwater runoff (Rg), snow melt amount (Rm), surface runoff (Rs), and snow depth water equivalent (SWE). These variables were based on Noah land surface model (Noah), version 2.1, which has been widely validated and applied in hydrological research for its robustness and high temporal resolution [46,47]. This version offers data at a spatial resolution of 0.25° and a temporal resolution of 3-hourly spanning from 1 January 2000 to the present (1 September 2022, used in this study). The study period for all variables used in the modeling was January 2003 to September 2022 to match the TWS data availability. Noah dataset can be accessed at https://daac.gsfc.nasa.gov/datasets/GLDAS_NOAH025_3H_2.1/summary (accessed on 10 January 2023). The overview of data used in this study is listed in

Table 1. Summary of the dataset used in this study.

Data Set	Variable Name	Acronyms	Unit	Date	Spatial Resolution	Temporal Resolution
GLDAS-CLSM (Version 2.2)	Terrestrial Water Storage	TWS	mm	Jan 2003–Sep 2022	0.25°	Daily
GLDAS-Noah (Version 2.1)	Baseflow-groundwater runoff	Rg	mm	Jan 2000–Sep 2022	0.25°	3 h
	Surface snow melt amount	Rm	mm
	Air temperature	T	K
	Precipitation	P	mm
	Surface runoff	Rs	mm
	Soil moisture	SM	mm
	Surface air pressure	SP	Pa
	Snow depth water equivalent	SWE	mm

.

To prepare the data for analysis, hydrological variables, including P, ET, Rg, Rs, and Rm, were first aggregated to a daily scale before multi-year averaging. For soil moisture, as Noah provides the SM data in four distinct layers, a weighted averaging method was applied to compute the total soil moisture. Regarding SWE, given its significant regional and seasonal variability, as well as its generally low magnitude, the months with highest coverage and values (January, February, November, and December) were selected for multi-year averaging to minimize potential biases.

2.3. Ensemble Machine Learning Framework

Extreme Gradient Boosting (XGBoost) has emerged as a powerful ensemble machine learning approach in hydrological studies for its superior performance in handling complex, non-linear relationships in Earth system data [48]. XGBoost is an ensemble learning method that sequentially trains weak learners (typically decision trees) to correct the errors of previous models, thereby improving overall prediction accuracy. It also incorporates regularization techniques to prevent overfitting [48,49,50,51]. Additionally, its computational efficiency is enhanced by parallel processing, out-of-core computation, GPU acceleration, and distributed training, making it particularly suitable for large-scale hydrological datasets. Moreover, regional-scale studies on TWS reconstruction have demonstrated that XGBoost achieves comparable or superior predictive performance compared to artificial neural networks and random forests while maintaining model interpretability [52,53]. Based on these advantages, XGBoost was selected for TWS simulation in this study.

The model used in this study was implemented using the native interface of the Python package “xgboost” (version 1.6.1) with Python 3.9.13. The dataset, comprising 94,703 samples, was randomly partitioned into training and testing sets at a 6:4 ratio, ensuring a similar distribution of data between the two sets. The best model, defined as the one with the minimum average root-mean-square error (RMSE), was selected based on 10-fold cross-validation conducted on the training set. For each fold, model hyperparameters were optimized by the stochastic hill-climbing algorithm, a local search method that sequentially adjusts each parameter within a predefined range. Starting from initial values and step sizes, the algorithm iteratively updates each hyperparameter to minimize RMSE, fixing the optimal value before proceeding to the next parameter. This process continues until a locally optimal hyperparameter combination is obtained. The hyperparameters considered, along with their optimal values, are summarized in

Table 2. Hyperparameters, search ranges, and final selected values used for XGBoost model optimization.

Hyperparameter	Search Range	Final Value
n_estimators	(800, 1200)	1200
learning_rate	(0.05, 0.2)	0.07
max_depth	(8, 12)	10
subsample	(0.6, 0.8)	0.6
reg_alpha	(1, 3)	2
reg_lambda	(1, 5)	3
gamma	(0.3, 0.8)	0.5

. The complete framework is illustrated in Figure 2.

2.4. Evaluation Metrics

The evaluation of the model was carried out using the combination of three metrics: coefficient of determination (R²), RMSE, and mean absolute percentage error (MAPE). R² is a common indicator that quantifies the goodness of fit, with higher R² values signifying better performance. RMSE measures the magnitude of prediction errors, with a smaller RMSE associated with a better performance. MAPE evaluates model performance from the perspective of relative error, which is particularly advantageous when comparing datasets with different units (e.g., simultaneously assessing TWS and discharge predictions). A lower MAPE indicates higher prediction accuracy, with MAPE < 15% generally indicating excellent model performance. R², RMSE, and MAPE are estimated as follows:

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2}}{{\displaystyle \sum _{i=1}^n{(y_i-\overline{y})}^2}}}$$

(1)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2}}$$

(2)

$$MAPE={\displaystyle \frac{100\%}{n}}{\displaystyle \sum _{i=1}^n\left|\frac{y_i-\overline{y}}{y_i}\right|}$$

(3)

where n typically denotes the sample size, $y_i$ represents the predicted value of the i-th sample point, ${\widehat{y}}_i$ denotes the observed value of the i-th sample point, and $\overline{y}$ is the average of observed values in the samples.

2.5. SHAP

In this study, Shapely additive explanations (SHAP) [54,55], which is grounded in the Shapley value concept from game theory, were employed to explain predictions of the machine learning model. SHAP assigns each feature an importance value representing its marginal contribution to the prediction, ensuring consistency, local accuracy, and fair attribution across features. In hydrological modeling, where physical processes are often complex, nonlinear, and interdependent, SHAP offers a robust framework for quantifying the relative influence of meteorological, climatic, and catchment-specific variables on model outputs. Moreover, compared to the interpretation method used in XGBoost (gain-based interpretation, GBI), SHAP is more effective in handling multicollinearity and offers a more comprehensive understanding of the results, thus facilitating subsequent decision-making processes [56]. The calculation of SHAP value can be simply represented as follows:

$$g\left(Z^{\prime }\right)={\phi }_0+{\displaystyle \sum _{j=1}^M{\phi }_i{Z}_i^{\prime }}$$

(4)

$${\phi }_i={\displaystyle \sum _{S\subseteq N/i}\frac{\left|S\right|!(M-|S|-1)}{M!}(f_x(S\cup i)-f_x(S))}$$

(5)

where Z′∈{0, 1}^M is a coalition vector representing the presence or absence of each feature, M is the total number of features, φ_i is the SHAP value of feature i, f(x) is the non-linear regression mapping of XGBoost, N is the sample set of feature variables with a dimension of M, and S is a subset extracted from N with a dimension of |S|. This ultimately reflects the difference in prediction values caused by the inclusion of the additional feature variable. Each data sample’s final prediction value, when inputted into the explanation model, corresponds to a SHAP value, which reflects the contribution of the corresponding driving factor’s importance [57].

3. Results

3.1. Model Performance

Figure 3 presents the model performance on the testing set from different perspectives. Generally, the XGBoost model achieved an R² of 0.98, a mean MAPE of 2.36%, and a mean RMSE of 22.88 mm, demonstrating strong predictive capability and overall model performance. Figure 3a,b display the multi-year average TWS of observations and predictions, respectively, where minimal discrepancies are observed. This suggests that the model can effectively capture the spatial patterns of TWS. Figure 3c,d compare observed and predicted TWS values in detail. A high degree of fit is evident, with most points closely aligning along the 1:1 line (y = x), even for extreme values. Moreover, the model achieved a median residual of −0.30 mm, a mean MAPE of 1.28%, and a mean RMSE of 12.08 mm, further demonstrating its strong predictive accuracy.

A more in-depth analysis of the spatial distribution of the MAPE was conducted to better understand model performance across different regions. Figure 4 illustrates the spatial distribution of the MAPE. The results demonstrate exceptional model accuracy, with approximately 70% of monitoring points achieving MAPE values below 2.5%, while over 99% of points fall within the 20% threshold. Furthermore, larger MAPE values are predominantly concentrated in the Yellow River Basin and arid northwestern regions, characterized by intensive anthropogenic activities such as irrigation withdrawals and reservoir operations. This spatial correspondence suggests that human-induced hydrological modifications may introduce additional uncertainty, leading to increased prediction errors in these sensitive zones. Overall, model predictions demonstrate strong agreement with observed TWS, as supported by the evaluation metrics, thus providing a robust foundation for subsequent SHAP analysis.

3.2. Main Drivers for TWS

Figure 5 illustrates the feature importance derived from the XGBoost model using GBI and SHAP methods. Both approaches consistently identify P and SM as the most important features, with SP also showing a relatively high influence. This strong agreement between the two methods enhances the credibility of the results. However, a divergence is observed in the evaluation of T, Rg, and Rs: the GBI method attributes relatively low importance to these variables, whereas the SHAP method suggests a greater contribution. This discrepancy may be attributed to the GBI method’s limited capacity to accurately account for multicollinearity among input features, which can obscure the true importance of correlated variables. These findings further underscore the necessity of employing the SHAP method for more robust and reliable interpretation of feature contributions.

To further explore the influence of factors on TWS, the dominant driving factors at the pixel level were identified and mapped in Figure 6. Results reveal a significant regional heterogeneity from southeast to northwest China. In the monsoon climate zones, P emerges as the primary driver for TWS. Precipitation-dominated and non-precipitation-dominated regions are roughly bounded by the 400 mm annual precipitation threshold, emphasizing the important role precipitation plays in the distribution of TWS in China. In the Qinghai–Tibetan Plateau, SM plays a dominant role, which can be attributed to the water retention capacity under complex terrain conditions. In the northwestern arid and semi-arid climate zones, TWS is mainly regulated by SP, Rg, and Rm, where precipitation is scarce and TWS recharge comes mainly from Rm and Rg.

3.3. Individual Impact of Driving Factors

Figure 7 illustrates the individual impact of each driving factor on TWS, with the interactions between factors removed (showing SHAP main effects). Values above the red line indicate a positive driving effect, while values below the red line indicate a negative driving effect. Generally, almost all factors show non-monotonic impacts on TWS. Among all factors, precipitation and soil moisture exhibit relatively strong positive monotonic effects. As P and SM increase, their positive contributions to TWS strengthen. However, when P and SM reach a certain threshold, their impacts on TWS weaken. This may indicate that at this point, the soil has reached its water-holding capacity, limiting further positive contribution from P. Similarly, the influence of Rs and Rg on TWS is non-monotonic and relatively strong at low values. However, as the values increase, the effect gradually weakens and stabilizes. Interestingly, when SP exceeds a certain threshold, its SHAP value turns sharply negative. In terms of SWE and Rm, they have certain impacts on TWS at low values, but as their values increase, their influence becomes almost negligible. This could be attributed to China’s extensive latitudinal range, where snow accumulation is limited to specific regions. Significant snowfall only occurs in certain high-latitude or high-altitude areas, where SWE exerts a pronounced positive effect on TWS.

4. Discussion

4.1. Spatial Distribution of Dominant Drivers of TWS in China

TWS is a key component of the global hydrological cycle, and its distribution reflects the availability of freshwater resources. Identifying the driving factors of the spatial TWS distribution helps to improve our understanding of the water cycle processes, while also aiding in the assessment and allocation of water resources [58]. Previous studies have demonstrated that climate factors (e.g., precipitation, temperature, and snowmelt) dominate TWS variability through both direct (e.g., recharge via rainfall) and indirect pathways (e.g., temperature-controlled evapotranspiration) [59,60].

According to the feature importance rankings in Figure 5, precipitation is the primary factor influencing the distribution of TWS. On average, China receives about 0.64 mm of daily precipitation, but there are substantial regional differences. Mean annual precipitation ranges from over 2000 mm in the southeastern coastal regions to less than 50 mm in the arid northwest. Influenced by the monsoon, the southern and eastern coastal regions experience higher daily precipitation, while the northern and western regions, particularly the northwest, have lower daily precipitation. This significant difference in daily precipitation is the fundamental reason why precipitation is the primary factor determining the distribution of TWS in China. However, given the large monsoon-influenced regions, precipitation is the primary source fueling TWS across much of the country. This also explains why precipitation dominates in the model outputs and exhibits a significant positive driving contribution in the SHAP value analysis.

However, Figure 7a shows that as precipitation increases, its positive contribution to TWS distribution decreases. Results reveal that a multi-year mean daily precipitation of 4 mm (approximately 1400 mm annually) corresponds to the peak contribution of precipitation to TWS. When precipitation exceeds this threshold, its contribution to TWS begins to decline sharply. This variation may result from the coupled effects of soil hydrological properties and vegetation ecological processes [61]. Figure 8b shows that this level of precipitation primarily occurs in the southeastern coastal regions and southern China. From a soil perspective, the saturated hydraulic conductivity (K_sat) of the main soil types in eastern China reaches the inflection point of infiltration efficiency when the daily precipitation is between 3.5 and 4.5 mm [62]. When precipitation exceeds this threshold, the decline in soil matric potential slows down the infiltration rate of water into the subsurface and groundwater. This leads to increased saturation-excess runoff, thereby reducing TWS. Concurrently, enhanced lateral flow diverts more water toward rivers or water bodies rather than being stored in the soil [42,53]. This precipitation level (approximately 1400 mm a⁻¹) closely matches the water demand of the maximum net primary productivity (NPP) zone (1200–1500 gC m⁻² a⁻¹) [63] in the eastern forests of China. At this point, canopy interception (about 15%), transpiration (about 35%), and soil water retention (about 40%) reach an optimal balance for water use [64]. However, further increases in precipitation disrupt this balance: canopy interception rises nonlinearly (increasing by 1.8% for every additional 100 mm), and increased respiratory losses reduce the water use efficiency of NPP, forming a negative feedback loop among precipitation, vegetation, and TWS [57,63,64,65].

Both importance metrics (GBI and SHAP) indicate soil moisture exerts the second largest importance to TWS. Figure 6 illustrates that the sample points where SM dominates are primarily located in the Tibetan Plateau region. This region receives relatively low annual precipitation (especially in the western and northern parts), making the soil’s water-holding capacity a key factor in water retention. The Qinghai-Tibet Plateau features highly variable terrain, with lithology dominated by clastic rocks and permafrost layers [66]. These conditions hinder the development of well-formed aquifers, making it difficult to establish stable groundwater systems. Under such circumstances, shallow soil moisture becomes the most stable and responsive component of TWS. Although the plateau region does experience seasonal snow and glacial cover, rising temperatures cause snowmelt and runoff to flow out of the region, failing to effectively contribute to regional TWS gains [67]. Instead, this outflow may intensify the declining trend of TWS. As a result, although SWE may have significant short-term contributions in certain areas, its overall spatial contribution remains limited. In terms of the arid northwest region, SP and Rg emerge as dominant contributors in the XGBoost model. With scarce precipitation (e.g., <100 mm a⁻¹ in the Tarim Basin) and intense evaporation, elevated SP suppresses surface evaporation, reducing moisture loss and indirectly preserving limited soil water and groundwater, thereby enhancing their contribution to TWS. Additionally, the region is influenced by the Siberian High, where SP variations directly modulate external moisture transport by regulating atmospheric circulation, making it a key predictor for TWS [68]. Meanwhile, minimal surface runoff and sustained groundwater recharge from mountain snowmelt (e.g., Tianshan and Kunlun Mountains) maintain stable TWS [58,69,70]. The high SHAP value of Rg underscores its critical role in sustaining water resources in this arid zone.

In this study, the effect of temperature on TWS is primarily negative via indirect pathways (Figure 7e). This aligns with the fundamental physical mechanism, whereby elevated temperatures increase atmospheric water demand. Firstly, enhanced evapotranspiration under warmer conditions depletes soil moisture and reduces groundwater recharge [71]. Secondly, the spatial gradients representing persistent elevated temperature states in our climatological analysis demonstrate significant coupling with hydrologically stressed regimes that systematically impair water storage recovery processes [60].

4.2. Sources of Uncertainty

This study employed GLDAS to construct XGBoost and SHAP interpretability; this process may introduce uncertainties at both the data and model structure levels. Firstly, the structure of the GLDAS land surface model has certain limitations. For example, the soil moisture variable output by the model only reflects the changes in near-surface soil moisture and does not account for the changes in deep groundwater [72]. Additionally, the impact of human factors on TWS has not been included in this study. According to statistics, irrigation accounts for about 70% of global freshwater withdrawals. As an important component of the water cycle, TWS is strongly influenced by irrigation [39,73]. In China, the total irrigated area is approximately 67 million hectares, distributed across various provinces and municipalities. Irrigation is a significant factor influencing changes in TWS, and since our model relies solely on climate data, the predictions may lack precision in heavily irrigated areas. Secondly, many reservoirs were built for irrigation, flood control, and other purposes in China. The large amount of water stored in these reservoirs may lead to long-term changes in TWS, making the GLDAS data insufficiently accurate for simulating TWS in the Chinese region [74]. Moreover, numerous water resource management initiatives, such as the South-to-North Water Diversion Project, the Dianchi Lake Restoration Project, and the renovation of the Grand Canal, also impact the spatial distribution of TWS.

XGBoost can effectively reduce uncertainty through ensemble learning and its relatively transparent model structure; however, like many powerful ML models, its internal complexity poses challenges for interpretation. XML techniques, while enhancing the model interpretability and enabling informed model adjustments, inevitably introduce additional uncertainties, as they rely on data-driven frameworks that are inherently limited by the underlying algorithms. Moreover, the additive nature of SHAP value attribution estimates the contribution of each feature individually during feature analysis [54], potentially overlooking the complex interdependencies and couplings that often characterize real-world physical processes. Such limitations can affect the completeness of the mechanistic interpretation derived from SHAP values. To address this, we carefully analyzed feature correlation and made appropriate trade-offs during feature selection and preprocessing.

5. Conclusions

Since the launch of GRACE satellites in 2002, TWS declines have been observed in China at a rate of approximately 2 mm a⁻¹. To uncover the main drivers and their mechanisms behind these TWS losses, XGBoost and SHAP methods were employed to identify the spatial pattern of TWS in China and quantified the independent relative contributions of various hydroclimatic factors. Our results reveal that the XGBoost model can efficiently reproduce TWS patterns across China, achieving an R² of 0.98 and RMSE of 22.88 mm. Furthermore, the SHAP method showed that, in general, precipitation and soil moisture contribute most significantly to TWS spatial variability, with predominantly positive impacts. While surface pressure, surface runoff, and temperature also affect TWS, their effects are largely negative.

Spatially, the dominant driving factors exhibit pronounced heterogeneity. Precipitation emerges as the primary driver in southern China and the North China Plain, while soil moisture plays a leading role across the Tibetan Plateau. In northwestern China, surface pressure, snow melt runoff, and groundwater runoff are identified as key determinants of TWS distribution. Furthermore, our results reveal that the influence effects of most factors are nonlinear: the importance of precipitation decreases beyond a certain threshold, while surface pressure transitions from a positive to a negative effect as its value exceeds a certain level. These findings enhance our understanding of the spatial patterns of TWS in China and offer valuable insights for more effective water resource management and allocation. Moreover, the interpretable machine learning model based on XGBoost and SHAP demonstrate strong adaptability in simulating TWS patterns, underscoring the applicability of explainable machine learning in hydrological research.