Identification of Driving Factors of Long-Term Terrestrial Water Storage Anomaly Trend Changes in the Yangtze River Basin Based on Multisource Data and Geographical Detector Method

Abstract

Terrestrial water storage anomaly (TWSA) plays a vital role in regulating the global water cycle and freshwater availability. Understanding the drivers behind long-term TWSA changes is critical, yet disentangling natural and anthropogenic influences remains challenging. This study employs the Geographical Detector method and multisource data to quantify the individual and interactive effects of multiple drivers on TWSA trends across the upper, middle, and lower reaches of the Yangtze River Basin (YRB). In the upper YRB, temperature, snow water equivalent, vegetation, precipitation, and reservoir storage are the primary contributors. In the middle YRB, precipitation, temperature, and soil moisture dominate. Although nighttime light (a proxy for urbanization) alone explains only 1.94% of the variation in this region, its interaction with precipitation increases explanatory power to 56.3%, highlighting a strong nonlinear effect. In the lower YRB, precipitation and runoff are the leading factors, while nighttime light again exhibits enhanced influence through interactions. These findings reveal the spatial heterogeneity and synergistic nature of TWSA drivers and underscore the need to consider both natural variability and human-induced processes when assessing long-term water storage dynamics. The results offer valuable insights for sustainable water resource management in the context of climate change and rapid urban development.

1. Introduction

As the third largest river basin in the world, the Yangtze River Basin (YRB) plays a vital role in China’s water resources supply and national development. The basin is characterized by high population density, intensive agricultural activities, and extensive hydraulic infrastructure such as dams, reservoirs, and irrigation systems [1]. Its social and economic development is highly dependent on abundant and stable water resources. Therefore, any significant changes in the availability or distribution of water resources in the YRB may not only impact regional sustainability but also pose implications for national and even global water security and ecological stability [2].

Over the past decades, variations in terrestrial water storage (TWS) in the YRB have been driven by the combined influences of natural factors (e.g., climate variability) and anthropogenic activities (e.g., land use change, urbanization, water conservancy operations) [3]. Investigating the impact of multiple variables on terrestrial water storage anomaly (TWSA) contributes to a deeper understanding of the basin’s hydrological cycle [4,5], while providing theoretical support for effective management and balanced allocation of water resources, which in turn improves utilization efficiency and reduces economic losses from hydrological hazards such as droughts and floods [6,7].

Since its deployment in March 2002, the Gravity Recovery and Climate Experiment (GRACE) mission has transformed the way we observe and track temporal variations in global water storage. GRACE, together with its successor mission GRACE Follow-On (GRACE-FO), has become an essential tool for quantifying basin-scale water mass fluctuations and detecting trends in TWSA with unprecedented precision [8,9]. Many researchers have investigated TWS variations in the YRB using GRACE-derived datasets. For example, Chao et al. [10] examined the components of TWS in the YRB by combining GRACE data, ground-based observations, and multi-source datasets, separating the contributions of glaciers, surface water, soil moisture, and groundwater. Wang et al. [5] developed a deep learning framework, RecNet, to reconstruct climate-driven TWSA changes from 1923 to 2022, achieving higher accuracy than conventional reconstruction approaches. Xie et al. [11] employed three machine learning techniques—namely multi-layer perceptron (MLP), long short-term memory (LSTM), and multiple linear regression (MLR)—to reproduce TWSA and proposed a new index capable of identifying extreme flood events at sub-monthly resolution.

Moreover, GRACE-derived TWSA data have enabled the development of various drought [12,13,14,15] and flood [16,17,18] indices, offering new approaches for assessing hydrological extremes in the YRB. These studies have provided valuable insights into the temporal variations in water storage. However, they often lack a quantitative identification of the driving mechanisms underlying long-term trends in TWSA.

Since the late 1990s, the YRB has undergone substantial land cover change [19], widespread vegetation greening driven by ecological restoration programs [20], and increasing climate variability [21,22,23]. Under the dual influences of natural and human-induced drivers, the trends of TWSA have shown significant spatial heterogeneity and temporal instability in recent decades. Therefore, identifying the dominant drivers and their spatial distribution is crucial for fine-scale water resource management and adaptation to global change–induced hydrological risks.

To systematically identify and quantify the relative influences and interactions of natural and anthropogenic factors on the long-term trends of TWSA in the YRB, this study employs the Geographical Detector (GeoDetector) model. GeoDetector is a spatial statistical tool specifically designed to uncover the driving forces behind spatial heterogeneity. It quantifies the explanatory power of multiple factors without requiring assumptions of linearity or data normality. The GeoDetector model has been successfully applied in various studies to quantitatively analyze the driving factors of environmental change, detect high-risk areas, and identify spatial variations [24,25,26,27].

Therefore, the objective of this study is to identify the key driving factors and their interactions that dominate the trends of TWSA changes in the YRB and to analyze their spatial differentiation. The findings will provide both data support and theoretical guidance for regional water resource management, contribute to adapting to complex hydrological conditions under global change, and offer a scientific basis for predicting water storage trends and formulating appropriate management strategies.

2. Materials and Methods

2.1. Study Area

The YRB ranks as the largest river system in China and the third longest worldwide. Its drainage basin covers approximately 1.8 million km², representing about 18.8% of the country’s total land area. The river originates from the main peak of the Tanggula Mountains on the Qinghai–Tibet Plateau. Its mainstream, together with numerous tributaries, traverses up to 19 provinces, with the trunk channel extending nearly 6300 km in length. The latitudinal and longitudinal coordinates of the basin span a range from 90°32′ to 121°56′ E and 24°28′ to 35°46′ N, respectively. The elevation span in the basin is vast, from −70 m to 6444 m. The basin’s terrain is structured in three distinct tiers, displaying a predominant gradient from higher elevations in the west to lower elevations in the east. Accordingly, it is conventionally divided into the upper, middle, and lower reaches. The upper reaches of the YRB, extending from the river’s source to Yichang in Hubei Province, are mainly composed of alpine mountains with an average altitude of about 5100 m. The middle reaches stretch from Yichang to Hukou, characterized by scattered alluvial plains and undulating low hills, with an average elevation of less than 200 m above sea level. The lower segment runs from Hukou to Chongming Island in the east, encompassing densely settled and highly urbanized regions along both the mainstream and its distributaries.

Due to the large area of the YRB, which spans more than 32 degrees of longitude and extends from the inland to the coast, the climate and hydrological characteristics of different regions vary considerably. Accordingly, the YRB is classified in this paper into three sub-basins arranged from west to east, namely the upper, middle, and lower portions. The overview of the study area is shown in Figure 1. The secondary basins in Figure 1 are numbered 1–12, respectively: (1) Jinsha River above Shigu, (2) Jinsha River below Shigu, (3) Mintuo River, (4) Jialing River, (5) Yibin to Yichang, (6) Wujiang River, (7) Hanjiang River, (8) Yichang to Hukou, (9) Dongting Lake system, (10) Poyang Lake system, (11) mainstream below Hukou, and (12) Taihu Lake system.

2.2. Data Sources

Selecting an appropriate GRACE product for the YRB is particularly important for ensuring the accuracy of subsequent analyses. Wang et al. [28] compared and evaluated TWSA derived from different GRACE products—spherical harmonics (SH), the Center for Space Research Mascon product (CSR-M), and the NASA Jet Propulsion Laboratory Mascon product (JPL-M)—using GPS data within the YRB. The results indicated that while the Mascon products capture more spatial detail in TWS distribution, the CSR-M product demonstrated higher consistency with GPS observations and better represented actual surface mass variations in the basin. Based on these findings, the CSR-M TWSA dataset is adopted in this study.

In addition, this study integrates a variety of hydrological, meteorological, reanalysis, and auxiliary datasets as potential driving factors. The selection of these factors was guided by data availability and reliability, as well as evidence from previous studies on hydrological processes [10]. A summary of the selected datasets is provided in

Table 1. Brief description of the data used for the driving factors.

Type	Driving Factors	Abbreviation	Temporal Resolution	Spatial Resolution	Data Source
Natural factors	Soil moisture storage	SMS	Monthly	0.1° × 0.1°	FLDAS Noah [29]
	Evapotranspiration	ET	Monthly	0.25° × 0.25°	GLEAM [30]
	Potential evapotranspiration	PET	Monthly	0.25° × 0.25°	GLEAM
	Precipitation	PRE	Daily	0.25° × 0.25°	CN05.1 [31]
	Temperature	TEM	Hourly	0.1° × 0.1°	ERA5 Land [32]
	Snow water equivalent	SWE	Hourly	0.1° × 0.1°	ERA5 Land
	Runoff	RO	Daily	0.25° × 0.25°	CNRD v1.0 [33]
	Normalized difference vegetation index	NDVI	Monthly	5 km × 5 km	A 5 km resolution dataset of monthly NDVI product of China (1982–2020) [34]
Anthropogenic factors	Nighttime-light	LIGHT	Monthly	100 m–1 km	A prolonged artificial nighttime-light dataset of China (1984–2020) [35]
	Reservoir water storage	RWS	Monthly	0.5° × 0.5°	WaterGAP Global Hydrology Model (WGHM) [36]

. The analysis is confined to the GRACE observation period, from 2002 to 2017.

2.3. Methodology

2.3.1. Data Pre-Processing

To ensure consistency among datasets with different spatial and temporal resolutions, all driving factor datasets were resampled to match the CSR-M mascon product. Specifically, the original datasets were bilinearly interpolated to a uniform spatial resolution of 0.25° × 0.25°. Furthermore, the temporal resolution of all datasets was aggregated to a monthly scale to align with the GRACE observation period. This preprocessing step ensured comparability of the datasets and minimized potential biases caused by inconsistent spatial and temporal resolutions.

2.3.2. Time Series Decomposition

The TWSA time series obtained from the GRACE(-FO) inversion contains a wide range of frequency variations, from long-term trends to monthly variations. These variations may be associated with different processes at different time scales. In order to identify the long-term trend term drivers in the YRB, this study decomposes the TWSA time series using least squares fitting with the following Equation.

$$\mathrm{TWSA}={\mathrm{TWSA}}_{\mathrm{trend}}+{\mathrm{TWSA}}_{\mathrm{inter}}+{\mathrm{TWSA}}_{\mathrm{seasonal}}+{\mathrm{TWSA}}_{\mathrm{subseas}}$$

(1)

GRACE TWSA raw signal was decomposed into linear trend term TWSA_trend, interannual term TWSA_inter, seasonal term TWSA_seasonal and sub-seasonal term TWSA_subseas. Similarly, other drivers were processed using the same methodology in extracting the linear trend term from 2002 to 2017 [37].

It should be noted that the GRACE and GRACE-FO satellite missions provide monthly TWSA estimates beginning in 2002, covering a relatively short observational period. While these data are well suited for detecting interannual to long-term trends, they may not fully capture decadal-scale variability or rare extreme events. Therefore, in this study we primarily focus on long-term basin-wide trends rather than short-term variability or individual extreme events.

2.3.3. The GeoDetector Model

The GeoDetector model is a spatial analysis tool grounded in the principle of spatial heterogeneity. Proposed by Wang et al. [38], this method is designed to identify and quantify the key drivers behind spatial phenomena. It is tailored to recognize spatial heterogeneity across strata and to investigate the mechanisms behind it. Originally applied in health risk assessment, the GeoDetector model has since been widely adopted across diverse disciplines, including ecology, meteorology, hydrology, economics, and the humanities.

The model comprises four core components: the factor detector, interaction detector, risk detector, and ecological detector. Suppose the dependent variable Y (e.g., normalized difference vegetation index, soil water storage) is affected by a set of independent variables X₁, X₂, X₃, X₄, …, X_n (e.g., solar radiation, surface reflectance, temperature, population density, climate zones, GDP, land use types, etc.).

Factor Detector: This module evaluates spatial variation in the dependent variable Y and assesses how strongly each independent variable X accounts for Y, represented by the q-value. The calculation is shown in Equation (2).

Risk Detector: This module evaluates whether there is a statistically significant difference in the mean values of the dependent variable between different subregions of an independent variable X.

Ecological Detector: It examines whether two independent variables, X_n and X_n+1, present meaningful variations in their influence over the spatial arrangement of Y.

$$\mathrm{q}=1-{\displaystyle \frac{{\displaystyle \sum _{\mathrm{h}=1}^{\mathrm{L}}{\mathrm{N}}_{\mathrm{h}}{\mathsf{\sigma }}_{\mathrm{h}}^2}}{\mathrm{N}{\mathsf{\sigma }}^2}}$$

(2)

In the above equation, h = 1, …, L represents the stratification, i.e., the categorization or partitioning of the dependent variable Y or the independent variable X. N_h and N denote the number of spatial units (cells) within stratum h and in the entire study area, respectively. σ_h² and σ² represent the variances of Y within stratum h and the entire study area, respectively, see

Table 2. Interaction judgment basis.

Judgment Basis	Interaction
$\mathrm{q}({\mathrm{X}}_1\cap {\mathrm{X}}_2)<\mathrm{M}\mathrm{i}\mathrm{n}(\mathrm{q}({\mathrm{X}}_1),\mathrm{q}({\mathrm{X}}_2))$	Non-linear weakening
$\mathrm{M}\mathrm{i}\mathrm{n}(\mathrm{q}({\mathrm{X}}_1),\mathrm{q}({\mathrm{X}}_2))<\mathrm{q}({\mathrm{X}}_1\cap {\mathrm{X}}_2)<\mathrm{M}\mathrm{a}\mathrm{x}(\mathrm{q}({\mathrm{X}}_1),\mathrm{q}({\mathrm{X}}_2))$	Single-factor nonlinear attenuation
$\mathrm{q}({\mathrm{X}}_1\cap {\mathrm{X}}_2){>\mathrm{M}\mathrm{a}\mathrm{x}\left(\mathrm{q}\right(\mathrm{X}}_1{),\mathrm{q}(\mathrm{X}}_2\left)\right)$	Two-factor interaction enhancement
$\mathrm{q}({\mathrm{X}}_1\cap {\mathrm{X}}_2)=\mathrm{q}({\mathrm{X}}_1)+\mathrm{q}({\mathrm{X}}_2)$	Mutual independence
$\mathrm{q}({\mathrm{X}}_1\cap {\mathrm{X}}_2)>\mathrm{q}({\mathrm{X}}_1)+\mathrm{q}({\mathrm{X}}_2)$	Non-linear enhancement

Notes: $\mathrm{q}({\mathrm{X}}_1)$ and $\mathrm{q}({\mathrm{X}}_2)$ are the q-values of the dependent variables ${\mathrm{X}}_1$ and ${\mathrm{X}}_2$, respectively, and $\mathrm{q}({\mathrm{X}}_1\cap {\mathrm{X}}_2)$ refers to the interaction of $\mathrm{q}({\mathrm{X}}_1)$ and $\mathrm{q}({\mathrm{X}}_2)$; $\mathrm{M}\mathrm{i}\mathrm{n}(\mathrm{q}({\mathrm{X}}_1),\mathrm{q}({\mathrm{X}}_2))$ is the minimum value of $\mathrm{q}({\mathrm{X}}_1)$ and $\mathrm{q}({\mathrm{X}}_2)$, and similarly, $\mathrm{M}\mathrm{a}\mathrm{x}(\mathrm{q}({\mathrm{X}}_1),\mathrm{q}({\mathrm{X}}_2))$ is the maximum value of $\mathrm{q}({\mathrm{X}}_1)$ and $\mathrm{q}({\mathrm{X}}_2)$, and $\mathrm{q}({\mathrm{X}}_1)+\mathrm{q}({\mathrm{X}}_2)$ is the sum of $\mathrm{q}({\mathrm{X}}_1)$ and $\mathrm{q}({\mathrm{X}}_2)$.

.

The GeoDetector model was applied using the R package “geodetector”, which is available at https://cran.r-project.org/web/packages/geodetector/vignettes/geodetector.html (accessed on 30 September 2025). This platform allows for efficient handling of large datasets and provides a robust environment for conducting spatial analysis of the TWSA components. All data processing and analysis were conducted using R 4.2.1 and Matlab R2024. The model was run on a standard desktop PC (Intel i7 processor, 32 GB RAM; Lenovo, Beijing, China), and the total CPU runtime was approximately 60 min per regional GeoDetector run.

In this study, we primarily apply the factor detector and the interaction detector modules of the GeoDetector model to identify and analyze the driving factors influencing the long-term trend component of TWSA in the YRB. These modules quantitatively evaluate the respective impacts of natural and anthropogenic factors on the long-term trend of TWSA, and determine whether the combined influence of any two factors results in an enhancement or weakening of their explanatory power.

Given the substantial differences in human activities, climate conditions, and environmental changes across the upper, middle, and lower reaches of the YRB, the analysis is conducted separately for these three subregions. By applying both the factor detector and interaction detector, this study aims to accurately identify the dominant driving forces and their interactions affecting the long-term trend in each subregion. The GeoDetector model detects spatial heterogeneity (i.e., anisotropy) in the data, attributes it to specific driving factors, and assesses the nature and strength of interactions among them.

3. Results

3.1. TWSA Changes in the YRB

Figure 2 shows spatial distribution of TWSA trend in the YRB during the study period. Overall, TWSA shows an increasing tendency, with particularly pronounced growth observed in the eastern upper reaches and the middle reaches, where the rate exceeds 5 mm/year. Nevertheless, the spatial distribution is heterogeneous. Since the 1960s, the Qinghai–Tibet Plateau has experienced a sustained warming trend, which has accelerated glacier ablation in the region. For instance, the melting of glaciers in the Tanggula Mountains has led to a reduction in terrestrial water storage in the central-western upper reaches of the YRB, at an approximate rate of −5 mm/year. At the same time, the substantial input of meltwater from snow and ice has effectively replenished TWS in other parts of the basin.

3.2. Identification of Driving Factors for the Long-Term Trend of TWSA in the Upper YRB

To quantitatively assess the explanatory power of various natural and anthropogenic factors on TWSA variations in the upper YRB, a Geodetector analysis was conducted. The results, expressed as Q values, are presented in Figure 3. A higher Q value indicates a stronger ability of the factor to explain the spatial differentiation of TWSA. Among the ten evaluated factors, TEM emerged as the most dominant driver, with a Q value of 0.3802, underscoring its significant influence on regional hydrological processes such as evapotranspiration, snowmelt, and glacier dynamics. This was followed closely by SWE, with a Q value of 0.3669, emphasizing the important contribution of cryospheric factors to TWS, especially in the basin’s upstream areas where snow accumulation and melting are common. NDVI ranked third (Q = 0.234), indicating that vegetation and its ecohydrological feedbacks substantially contribute to TWSA variability. Climatic inputs such as PRE and PET had comparatively lower Q values of 0.1671 and 0.0773, respectively, indicating their indirect or regionally variable influence.

Anthropogenic factors exhibited moderate explanatory power. RWS had a Q value of 0.1531, reflecting the role of managed water infrastructure in regulating basin-scale hydrology. In contrast, LIGHT, a proxy for human settlement and activity intensity, showed relatively low influence (Q = 0.092).

The single-factor analysis based on the GeoDetector results suggests that TEM and SWE are the dominant driving factors affecting the TWSA trend component in the upper reaches. In contrast, NDVI, PRE, and RWS serve as moderately influential factors, while the remaining driving variables exhibit relatively weak impacts.

The ablation of glaciers induced by warming in the Tanggula Mountains has led to a reduction in terrestrial water storage in certain parts of the upper YRB at a rate of approximately −5 cm/year. Simultaneously, the melting of snow and ice has increased terrestrial water reserves in other regions of the basin. Overall, glacier ablation in the Tanggula Mountains has made a non-negligible contribution to the TWS in the YRB.

As a critical ecological factor, the spatial distribution of TWSA plays an essential role in shaping vegetation patterns. Conversely, vegetation significantly influences regional water reserves by regulating PRE, RO, ET, and SMS through water absorption, transpiration, interception, infiltration, and woodland storage.

Since the Three Gorges Dam became operational in 2003, it has undergone three reservoir impoundment phases. By 2010, it reached its final stage, with a normal storage level of 175 m. The storage and discharge operations of the Three Gorges Reservoir have also become important anthropogenic factors contributing to the long-term changes in TWSA in the upper reaches of the YRB.

The interaction between different driving factors may exert an even greater influence on TWSA trends. The interaction detector in GeoDetector facilitates the analysis of interactions by examining the q-values of each factor in relation to the q-values derived from factor pairs. In practice, the long-term trend of TWSA in the YRB is shaped by the combined effects of multiple factors.

The results of interaction analysis among the driving variables are presented in Figure 4. As shown, all pairwise interactions between driving factors yield stronger explanatory power for the upstream TWSA trend compared to individual factors. In particular, TEM and SWE exhibit markedly stronger impacts when interacting with other variables, highlighting their pivotal role in driving long-term TWSA changes in the upper YRB.

The types of interactions between individual driving factors, as illustrated in Figure 5, reveal that NDVI exhibits a two-factor enhancement effect when interacting with all other variables, except for PET, with which it shows a nonlinear enhancement. As a key ecological indicator of vegetation growth and development, the distribution of water storage plays a decisive role in vegetation dynamics. In turn, vegetation regulates regional water processes—such as PRE, ET, RO, and SMS—through water absorption, transpiration, interception, infiltration, and subsurface storage in the forest floor. Our analysis indicates that the interaction between NDVI and PET has a significant impact (non-linear enhancement effect) on the TWSA trend in these specific regions.

Additionally, some other driving factors also show nonlinear enhancement effects when interacting. In particular, the interactions among RO, SWE, and RWS belong to this category. Rising temperatures accelerate glacier ablation, which increases SWE and, in turn, strongly influences both RO and TWSA. As glacier melting intensifies, RO increases—until the glacier mass can no longer sustain the loss. At this point, runoff reaches its peak and then begins to decline as the glacier reserves are gradually depleted.

This process clearly demonstrates a nonlinear enhancement effect: at first, the interaction between SWE and RO amplifies the impact on TWSA because additional meltwater enters the hydrological system. However, once glacier mass falls below a critical threshold, the interaction weakens, causing RO to drop sharply and changing its contribution to TWSA. These dynamics illustrate the complex feedbacks between glaciers and hydrology in the upper basin.

Based on the above analysis, it is evident that TEM and SWE exhibit consistently higher explanatory power when interacting with other driving factors. These auxiliary variables—such as RO, ET, PET, and RWS—further enhance the impact of the dominant factors on the trend of upstream TWSA through synergistic interactions. This indicates that the dominant factors not only have strong individual effects but also exhibit amplified influence when coupled with supporting environmental variables.

3.3. Identification of Driving Factors for the Long-Term Trend of TWSA in the Middle YRB

Using the Geodetector method, the key driving factors influencing the spatial variation of TWSA in the middle YRB were identified and ranked based on their Q values (Figure 6). A larger Q value signifies a greater capacity of the corresponding factor to explain variations in the spatial heterogeneity of TWSA.

The results show that PRE is the most dominant factor, with a Q value of 0.5369, indicating that water storage variations in this region are primarily governed by direct hydrological input. This is followed by TEM and SMS, exhibiting Q values of 0.5174 and 0.5065, respectively. They significantly influence evapotranspiration, infiltration, and land–atmosphere water exchanges. RO, with a Q value of 0.3937, also exerts a significant influence, highlighting its intermediary role in connecting precipitation input with changes in basin water storage.

In contrast, other factors such as NDVI (Q = 0.1258), SWE (Q = 0.1109), and RWS (Q = 0.1012) have relatively lower explanatory power in the middle reaches. This may reflect the limited cryospheric processes in this region and the localized or seasonal influence of reservoir regulation. Anthropogenic influences are minimal in this area. LIGHT—used as a proxy for human activity—shows the weakest explanatory strength (Q = 0.0194), indicating a limited spatial impact of human presence on TWSA variation. PET and actual ET also have low Q values (0.0477 and 0.0232, respectively), possibly due to their indirect effects or interactions with more dominant climatic drivers.

Figure 7 presents the outcomes of the interaction analysis for the key influencing variables in the middle YRB. Notably, the interaction between TEM and SMS yields a q-value as high as 71.1%, indicating strong combined explanatory power for TWSA trends. This finding suggests that TEM changes substantially influence the dynamics of SMS, and together these factors exert a significant impact on water storage variability.

Similarly, the interaction between RO and TEM exhibits a high explanatory power of 66.3%. This high q-value reflects their joint contribution to explaining TWSA variation rather than a direct correlation between the two factors. The recent rise in TEM across the YRB has been linked to more frequent drought events, which alter RO processes and, in combination with TEM, further amplify the variability of water storage.

According to the interaction types illustrated in Figure 8, three factors—ET, PET, and LIGHT—demonstrate notable nonlinear enhancement effects when interacting with other driving factors. Specifically, interactions where LIGHT combines with PRE, TEM, and SMS exhibit strong explanatory power, with Q-value reaching 56.3%, 53.9%, and 53.3%, respectively. These results emphasize the increasing impact of human activities, as reflected by LIGHT, on hydrological processes in the middle YRB. The influence becomes particularly pronounced when human-induced factors interact synergistically with key climatic and hydrological variables. As an indicator of human activity, LIGHT data can directly reflect changes in the social environment resulting from urbanization and other anthropogenic processes. During urban development, continuous city expansion alters the underlying surface characteristics, affecting the exchange of water and heat between the land surface and the atmosphere. Combined with the release of anthropogenic heat, this leads to the formation of urban heat island and dry island effects, which in turn contribute to increased high-temperature events and more frequent droughts across the basin.

Moreover, due to the obstructive effects of heat islands and tall buildings, localized heavy rainfall tends to increase within urban areas. Existing studies have confirmed that urbanization exerts a measurable influence on regional climate dynamics.

The middle reaches of the YRB contain a mega-urban agglomeration centered around Wuhan, including the Wuhan City Circle, the Changsha-Zhuzhou-Xiangtan Urban Agglomeration, and the Poyang Lake Urban Agglomeration. After interacting with other environmental driving factors, urbanization processes in some parts of the middle reaches have further amplified their influence on the long-term trend of TWSA in this region.

3.4. Identification of Driving Factors for the Long-Term Trend of TWSA in the Lower YRB

Figure 9 shows the Q values of each driving factor in the lower YRB, the results clearly indicate that PRE is the most dominant factor in the lower YRB, with a Q value of 0.575. This underscores the primary role of direct hydrometeorological input in controlling terrestrial water storage, particularly in this lowland region where surface water dynamics are highly responsive to rainfall variability. RO ranks second (Q = 0.3191), reflecting its critical intermediary role in transporting water and linking precipitation to changes in water storage. This is followed by SMS (Q = 0.2547), which highlights its importance in regulating infiltration, evapotranspiration, and land surface hydrology. Notably, LIGHT ranks as the fourth most influential factor with a Q value of 0.2206, suggesting a notable human impact on water storage in the lower reaches—likely associated with urbanization, industrial development, and infrastructure concentration. This is a marked difference from the upstream and midstream results, where human activities played a less dominant role.

Climatic variables such as TEM (Q = 0.2155) and PET (Q = 0.2028) also show moderate influence. RWS, although highly managed in this region, ranks slightly lower (Q = 0.1998), possibly due to spatial averaging effects or the regulatory function of major reservoirs buffering TWSA fluctuations. NDVI, SWE, and ET exhibit relatively weak explanatory power, with Q values of 0.1132, 0.1047, and 0.0888, respectively. This is consistent with the region’s subtropical humid climate, low cryospheric influence, and dense human-altered landscapes.

The magnitude of precipitation change is greater in the middle and lower reaches of the YRB, making PRE the top-ranked driver of TWSA trend change in both regions. In contrast, the influence of RO increases progressively from upstream to downstream: RO in the upper reaches is relatively limited, increases in the middle reaches, and becomes most dynamic in the lower reaches.

In the lower basin, precipitation is abundant and primarily occurs in the form of liquid rainfall, which rapidly drains into river systems and ultimately flows into the East China Sea. Additionally, due to glacial melting caused by global warming in the headwater regions of the YRB, meltwater contributes to terrestrial water reserves across upstream and midstream areas. These reserves, in turn, feed downstream runoff, further influencing water availability in the lower basin.

Figure 10 presents the results of interaction analysis between driving factors in the lower YRB. The results reveal that all two-factor combinations enhance the explanatory power for downstream TWSA trend changes. Particularly, PRE exhibits a q-value exceeding 60% when combined with any other factor.

For instance, although the NDVI has a limited individual influence (q = 11.3%), its interaction with precipitation yields a significantly enhanced influence of 67.7%. This highlights vegetation’s role in intercepting water flow, storing moisture, and reducing river and lake runoff during high-precipitation periods, thus playing a role in flood mitigation. At the same time, increased precipitation also stimulates vegetation growth, reinforcing this relationship.

Similarly, the SWE as a standalone factor shows a relatively weak influence (q = 10.4%), but when combined with precipitation, the influence increases to 70.5%. Historical records show intense snowfall events in the lower YRB during January 2008, November 2009, February 2013, and January 2018. Given the lower YRB’s geographical characteristics—proximity to the estuary and flat plains—snowmelt rapidly enters river systems alongside rainfall, contributing to runoff that eventually reaches the estuary.

In the lower YRB, the combined effect of PRE and LIGHT demonstrates considerable explanatory strength, with a q-value of 64.8%. As previously discussed, LIGHT data effectively reflect the level of urbanization and intensity of human activities. The study by Wang et al. [39] further supports this finding, demonstrating that human activities contribute to a decline in annual precipitation across the YRB, with the most pronounced reductions occurring in the central-upper and southeastern-lower reaches, where anthropogenic pressures are more intense.

These processes—urban expansion, land use change, and infrastructure development—alter the underlying surface conditions and atmospheric dynamics, thereby exacerbating the long-term downward trend in precipitation and amplifying hydrological variability. All of these effects ultimately influence the long-term trend of TWSA in the region.

Figure 11 further indicates that nonlinear enhancement dominates the interaction types of interaction between driving factors in the lower YRB. This implies that the combined impact of human and natural factors on TWSA trends is not simply additive but synergistic, often resulting in amplified effects that cannot be ignored in regional water resource assessments and planning.

4. Discussion

4.1. Comparison with Results from Other Methods

In order to evaluate the performance of the GeoDetector model and examine the relative importance of the driving factors, we also applied the method proposed by Lindemann, Merenda, and Gold (LMG) [40]. This method estimates the relative contribution of each driving factor to R², explicitly considering the sequence in which the factors are introduced into the model.

In the upper YRB, the LMG-based results (Figure 12) indicate that TEM, NDVI, SWE, and RWS are the dominant factors influencing the TWSA trend. These findings are largely consistent with the GeoDetector analysis, underscoring the strong role of cryospheric processes, vegetation dynamics, and reservoir regulation in shaping TWSA changes in this region.

Figure 13 presents the contribution rates of various factors, as calculated using the LMG method, for the middle YRB. PRE, SMS, TEM, and RO together explain over 80% of the total contribution to TWSA trend variations. Importantly, the four dominant factors identified by the LMG method are consistent with the top four most influential single factors revealed by the GeoDetector model, suggesting strong agreement between the two approaches in this region.

In the lower YRB, the LMG results (Figure 14) show that PRE has the highest contribution, reaching 49.56%. Although the ranking of individual factors differs slightly from the results of the GeoDetector analysis, the overall discrepancies are minor. Both methods highlight PRE and RO as the primary drivers in this region, with anthropogenic influences playing a secondary role.

Overall, the LMG method provides a complementary perspective to the GeoDetector model. While GeoDetector effectively identifies spatial stratified heterogeneity and interaction effects, LMG offers a clear quantification of the relative contribution of each factor to model explanatory power. The consistency observed between the two approaches across the three sub-basins enhances confidence in the robustness of the results.

4.2. Limitations of Driver Selection and Analysis

Over the past five decades, climate change and anthropogenic activities have jointly driven the long-term trends of TWSA in the YRB. Among these, anthropogenic activities have exerted particularly strong influences on basin-wide water resource dynamics. However, due to limitations in data availability, this study incorporates only two indicators of human activity: LIGHT and RWS. It should be noted that the RWS data used in this study are derived from model-simulated estimates provided by the WGHM, which may differ from observational records. This limitation not only affects the accuracy of the results but also underscores the need for future studies to integrate higher-precision or in situ socio-economic and hydrological datasets in order to more comprehensively capture the anthropogenic drivers of TWSA variability.

Since the 1950s, more than 50,000 dams have been constructed across the YRB to support water supply, hydroelectric power generation, flood control, and irrigation. Figure 15 illustrates the change in total reservoir storage in the YRB since 1997, as reported by the Yangtze River Water Resources Commission. The figure clearly shows a significant upward trend in reservoir volume after 2002, highlighting the growing role of dam storage in altering regional hydrology.

Looking ahead, future studies should aim to incorporate higher-precision or in situ measured reservoir storage datasets to more accurately assess the impact of dam operations on the long-term trends of TWSA in the YRB.

In addition to the factors considered in this study, other anthropogenic activities—such as groundwater extraction, agricultural irrigation, and the South–North Water Diversion Project—also significantly influence TWSA changes and should be incorporated in future analyses. The impacts of human activities can be further quantified and characterized using variables such as population growth rate, gross domestic product (GDP), land use change, and shifts in industrial and agricultural development patterns. In future work, we plan to integrate more comprehensive socio-economic and hydrological datasets (e.g., groundwater extraction statistics, irrigation data, GDP, land use change, and water diversion project records) to better capture the anthropogenic drivers of TWSA variability.

Moreover, the mechanisms through which meteorological changes and human activities influence TWSA trends require deeper investigation. In particular, the interaction effects among different driving factors should be further explored, as these interactions may lead to nonlinear or synergistic outcomes that are not evident through single-factor analysis. Therefore, future studies should strive to expand the set of driving variables, with special attention to the interaction mechanisms among natural and anthropogenic factors.

Another important limitation is related to the coarse spatial resolution of GRACE-derived TWSA data (~300 km). This resolution may smooth out local-scale hydrological variations and increase uncertainty in attributing specific drivers in a heterogeneous basin such as the YRB. To mitigate this limitation, we adopted a sub-basin partitioning strategy and incorporated higher-resolution auxiliary datasets (e.g., precipitation, soil moisture, vegetation indices) to strengthen the robustness of the attribution results. Nevertheless, the potential loss of local details remains an inherent limitation that should be considered when interpreting the findings.

In this study, we conducted a basin-partitioned analysis to quantitatively identify the dominant drivers of TWSA long-term trend changes across the upper, middle, and lower reaches of the YRB. However, the rationality of using administrative or hydrological zoning as the basis for driver identification remains open to discussion. Subsequent research could consider adopting alternative zoning strategies based on climate classifications, such as humid versus arid regions, or ecological zones, to enhance the spatial sensitivity and scientific rigor of the analysis.

4.3. Limitations of the GeoDetector Model

The GeoDetector model can not only assess the influence of a single driving factor on the trend change of TWSA in the YRB, but also quantify the impact of two-factor interactions on the same trend. However, it is important to note that the influence of a single factor, as calculated by GeoDetector model, reflects the explanatory power of that factor in isolation, i.e., without considering the effects of other variables.

Thus, the q-statistic for each factor does not represent its relative importance in a multi-factorial system. In other words, the sum of all individual q-values does not equal 100%, and cannot be directly interpreted as a proportional contribution across all variables. Consequently, although the GeoDetector model offers valuable insights into spatially stratified heterogeneity and factor interactions, it does not directly indicate the relative contribution of an individual factor compared with the combined effects of all variables influencing TWSA trend changes.

5. Conclusions

This paper investigates how various natural and human-induced drivers affect long-term TWSA trends in the YRB using the GeoDetector model. The study examines both the individual contribution of each driver to TWSA and the interactive effects between pairs of drivers, offering a thorough understanding of the processes governing TWSA changes across different basin regions. The key findings are outlined below:

(1) In the upper reaches of the YRB, the most influential factors affecting the TWSA trend include TEM, SWE, NDVI, and RWS. In this alpine region, glaciers and snowpack serve as key components of the terrestrial water system. Rising temperatures under the influence of global warming have accelerated glacier melting, especially at the headwaters of the basin, contributing substantial meltwater to rivers and altering long-term TWSA trends. At the same time, ecological conservation and reforestation policies have significantly improved vegetation cover, enhancing the ability of vegetation to intercept precipitation, retain soil moisture, and reduce runoff. These biophysical processes further regulate water storage changes over time. Additionally, the regulation and operation of large reservoirs such as the Three Gorges Dam, especially in conjunction with climatic factors, also exert considerable influence on the regional hydrological balance and long-term water storage dynamics.

(2) Within the central section of the YRB, the primary drivers of TWSA trends are PRE, SMS, TEM, and RO. Notably, GeoDetector results indicate that combined interactions among factors frequently yield stronger explanatory strength than individual variables alone, with most interaction effects surpassing 50%. In particular, the interactions involving temperature and precipitation demonstrate a significant compound influence on TWSA variability. Interestingly, while nighttime LIGHT alone exhibits only a marginal influence (1.94%), its interactions with precipitation and temperature result in influence rates of 56.3% and 56.9%, respectively. This suggests that urbanization, as reflected by nighttime light intensity, may have an indirect influence on regional water storage patterns, potentially through interactions with climatic factors. However, the exact mechanisms require further investigation. The presence of major urban agglomerations (e.g., the Wuhan City Circle and Poyang Lake urban cluster) in this region may intensify these effects.

(3) In the lower reaches of the YRB, precipitation and runoff are the most significant drivers of TWSA trend changes. Both variables exhibit higher magnitudes than the basin-wide average and follow a spatial gradient, decreasing from coastal areas inland. The lower basin is primarily composed of low-lying plains, and its proximity to the estuary leads to a highly dynamic hydrological response to rainfall. The abundance of liquid precipitation in this region means that runoff is rapidly generated and flows into the river system, eventually draining into the East China Sea. Furthermore, interaction effects between precipitation and other drivers in this region are especially prominent, further reinforcing the dominant role of compound hydrometeorological processes. For instance, precipitation interacting with NDVI or SWE significantly amplifies its effect on TWSA trends, underscoring the importance of both vegetation processes and snowmelt dynamics in shaping water storage evolution in the lower basin.

Collectively, these findings demonstrate that the long-term trends of TWSA in the YRB are governed by a combination of spatially variable climatic, hydrological, and anthropogenic factors, with synergistic interactions playing a particularly critical role. The region-specific insights gained from this study offer valuable guidance for water resource management and hydrological modeling under the dual pressures of climate change and human disturbance. The integrated application of GeoDetector in this study provides a novel framework for disentangling multi-source drivers of water storage dynamics in large river basins, offering valuable guidance for adaptive water resource management in the context of global change.