Differentiating system performance and coordination in the water–energy–food nexus: evidence from the Xiangjiang river basin, China

1 Introduction

Global environmental change, resource scarcity, and rising food demand are reshaping regional sustainable development (Saleem et al., 2025). Climate change exacerbates water scarcity and the uneven spatial–temporal distribution of water resources, while growing energy demand and low-carbon transition pressures create structural tensions in energy systems (Wang et al., 2025). In particular, energy-intensive manufacturing sectors such as the textile industry face mounting pressure to decarbonize under the dual control of carbon emissions and carbon-trading prices (Yan et al., 2026). Concurrently, population growth and dietary shifts intensify challenges to global food security (Ghosh et al., 2024). These pressures are interlinked through complex resource flows and policy feedback, forming a highly interdependent system. The water–energy–food (WEF) nexus provides an integrated framework to capture these interdependencies: water supports energy production and irrigation, energy drives water supply and food processing, and land–water constraints influence food production (Al-Saidi and Elagib, 2017). River basins, as spatial units where natural processes and socio-economic activities converge, are critical domains for intensive WEF interactions and complex land-use decisions. Basin-scale WEF analysis thus enables coordinated assessment of hydrology, energy, and agriculture, informing sustainable land-use planning and ecological security management (Zheng et al., 2025). For instance, Zheng et al. (2025) combined a Super-SBM efficiency model, a coupling coordination model, and PLS-SEM to map spatially differentiated drivers of WEF coordination across 53 cities of the Yellow River Basin, underscoring the value of pairing coordination assessment with interpretable, city-scale driver analysis. Beyond this macro-scale resource perspective, environmental sustainability has also been examined at the corporate and sectoral level: recent evidence shows that water pollution can erode corporate ESG performance (Ma et al., 2026a), while socio-cultural factors such as indigenous religious traditions shape corporate greenwashing behavior in China (Ma et al., 2026b). These micro-level findings complement basin-scale analyses by linking regional resource governance to the environmental responsibility of economic actors.

Over the past decade, research on the WEF nexus has evolved from conceptual frameworks to empirical analyses of complex systems, focusing on resource security assessment, scenario analysis, optimization, and coordinated policy design. Three main lines have emerged: (1) macro-scale assessments of system security and vulnerability, including Hoff (2011) who introduced the WEF nexus framework emphasizing water availability and cross-sectoral coordination, Biggs et al. (2012) who identified synergies and trade-offs for sustainable livelihoods, Simpson and Jewitt (2019) who developed a global WEF nexus index, and Sun et al. (2021) who applied system dynamics (SD) modeling in the Yellow River Basin to explore multi-scenario system evolution; (2) studies on the transmission of resource constraints and inter-sectoral flows, exemplified by Daher and Mohtar (2018) with the WEF Nexus Tool 2.0 and Owen et al. (2018) tracing United Kingdom supply chain interactions via multi-regional input–output analysis; and (3) analyses of synergies and trade-offs under combined climate and human pressures, including Howells et al. (2013) using multi-model coupling, and recent Chinese studies integrating SD modeling, Copula functions, deep reinforcement learning, coupling coordination models, and geographically weighted regression to evaluate system resilience, policy impacts, and spatial heterogeneity (Zhang et al., 2024). Recently, coupling coordination models and spatiotemporal analyses have become widely applied for assessing coordinated development of WEF subsystems and identifying key driving factors.

At the methodological level, WEF research exhibits a distinct characteristic of multidisciplinary integration. System dynamics, which depicts the evolution of complex system behaviors through causal feedback loops, has been widely applied in studies. Input–output analysis reveals the implicit resource flows and interdependencies among sectors, as demonstrated in the research by Zhang et al. (2022). The coupled application of multi-scenario simulation and optimization algorithms, as in Howells et al. (2013), provides quantitative decision support for coordinated resource allocation. In recent years, coupling coordination models and spatiotemporal analysis methods have been increasingly employed to assess the coordinated development of water, energy, and food subsystems. These approaches, by constructing comprehensive evaluation indicator systems and measuring the strength of interactions and the degree of coordinated development among subsystems, offer effective tools for identifying regional differences and evolutionary trends. In parallel, advances in statistical learning and dimensionality-reduction techniques such as PCA and scaled-PCA methods for extracting informative signals from heterogeneous, high-dimensional data have further enriched the analytical toolkit for capturing complex variable interactions (Zhang et al., 2026). This provides a solid methodological foundation for the approach adopted in this study.

Although the water, energy, and food systems are inherently interconnected through resource flows and mutual dependencies, their interactions are highly nonlinear, asymmetric, and spatially heterogeneous. Water supports both energy production and agricultural irrigation, while energy is required for water extraction, treatment, and food processing, and food production depends on both water availability and energy inputs. These interdependencies are further complicated by regional differences in resource endowments, economic development, and industrial structure, leading to substantial spatial variation in the strength and direction of WEF interactions across different sub-regions. Therefore, analyzing each subsystem in isolation is insufficient to capture the systemic behavior of coupled resource systems. This raises an important research question: how the WEF subsystems interact and co-evolve under spatial heterogeneity, and what factors drive their coordination and imbalance.

To address the above gaps, this study takes the Xiangjiang River Basin as a typical case to systematically examine the spatiotemporal evolution, spatial dependence, and driving mechanisms of the WEF system at the prefecture-level scale. Specifically, this study constructs a comprehensive evaluation framework integrating the entropy weight–TOPSIS method and the coupling coordination degree model to assess both the overall system level and the coordination among water, energy, and food subsystems. Furthermore, spatial autocorrelation analysis is employed to identify spatial patterns and clustering characteristics, while the XGBoost–SHAP framework is introduced to quantitatively reveal the nonlinear effects, relative importance, and heterogeneous influences of multiple driving factors.

Compared with existing studies, this research makes three main contributions. First, it shifts the analytical scale from macro-level assessments to the river basin’s internal units, capturing the spatial differentiation between upstream, midstream, and downstream regions and providing more refined insights for basin-scale governance (Simpson and Jewitt, 2019). Second, it explicitly links the evolution of the WEF system with land-use structure and regional transformation processes, building on evidence that land-transfer reforms and green-productivity synergies can jointly advance sustainable agriculture (Fan et al., 2026), thereby enhancing the explanatory power of spatial–temporal analysis (Liu et al., 2019). Third, by integrating machine learning with interpretable methods, this study uncovers the differential driving mechanisms of the composite WEF system level and the coupling coordination degree, highlighting a distinction between system development and system coordination that is not explicitly addressed in comparable basin-scale WEF studies, which typically model the drivers of coordination alone (Yang et al., 2025).

2 Materials and methods

2.1 Research area

The Xiangjiang River Basin, located in the south-central part of China (Figure 1), constitutes a key sub-basin of the middle reaches of the Yangtze River and serves as the socio-economic core of Hunan Province. The basin spans major prefecture-level cities including Changsha, Zhuzhou, Xiangtan, Hengyang, Shaoyang, Yueyang, Chenzhou, Yongzhou, and Loudi. Topographically, it exhibits a clear gradient from south to north, characterized by mountainous and hilly terrains in the upper reaches, transitioning into undulating hills and alluvial plains in the middle and lower reaches. Elevation varies significantly, with higher altitudes concentrated in the southern and western regions and lower elevations in the northern plains (Deng et al., 2025). The basin lies within a subtropical monsoon climate zone, featuring abundant annual precipitation, pronounced seasonal variability, and frequent occurrences of floods and droughts, which jointly shape the regional water resource dynamics. Land use is dominated by a mix of cultivated land, forest ecosystems, and expanding urban built-up areas, reflecting both its role as an important grain-producing region and a rapidly urbanizing economic corridor. Administratively, the study adopts prefecture-level cities as the basic spatial units, encompassing substantial variation in population size, economic development (GDP), and agricultural production capacity. This heterogeneity provides an appropriate context for analyzing the spatial–temporal evolution and interactions of the Water–Energy–Food (WEF) system within the basin.

2.2 Data sources

This study integrates multi-source datasets to construct the Water–Energy–Food (WEF) evaluation system and to support the subsequent spatial and driving-factor analyses. The data form a balanced panel covering a continuous 21-year period (2002–2022) for the nine prefecture-level cities of the Xiangjiang River Basin (Changsha, Zhuzhou, Xiangtan, Hengyang, Shaoyang, Yueyang, Chenzhou, Yongzhou, and Loudi), yielding 189 city-year observations (9 cities × 21 years) for each indicator.

Water resource data, including total water inflow, surface and groundwater resources, total water withdrawal, and sectoral water use (agricultural, industrial, domestic, and ecological), were obtained from the Hunan Water Resources Bulletin and the China Water Resources Statistical Yearbook. Energy-related data, covering total energy production and consumption, energy structure, and energy-efficiency indicators such as energy consumption per unit GDP, were drawn mainly from the China Energy Statistical Yearbook and the Hunan Statistical Yearbook. Food (grain) data, including total grain output, crop yield per unit area, effective irrigation area, agricultural inputs, and rural electricity consumption, were compiled from the China Rural Statistical Yearbook and the relevant provincial yearbooks. City-level socio-economic indicators (per capita GDP, industrial structure, urbanization rate, energy intensity, and fiscal expenditure) were obtained from the Hunan Statistical Yearbook and the statistical yearbooks and statistical communiqués of the respective cities, whereas land-use composition and topographic characteristics were derived from remote-sensing products and publicly available geospatial datasets.

To ensure data reliability and reproducibility, three quality-control procedures were applied. First, each indicator was cross-checked across multiple official sources (national yearbooks, provincial yearbooks, and sectoral bulletins); where discrepancies arose, the value from the most authoritative and temporally consistent source was retained. Second, statistical-caliber consistency was verified across years to account for periodic revisions in administrative boundaries and statistical definitions, and series affected by such revisions were adjusted to a common basis. Third, the completeness of each series was assessed prior to analysis; overall, missing values accounted for approximately 21% of all observations and were concentrated in the early years (2002–2005) and in the ecological water-use and energy-structure indicators.

Missing values were handled as follows. For gaps occurring between two observed years, linear interpolation was applied; for missing values at the beginning or end of a series, extrapolation based on the average annual growth rate of the nearest available period was used. No single indicator exceeded a missing-data threshold of 50%, and all indicators listed in Table 1 were therefore retained. All imputed values were subsequently checked against provincial-level aggregates for plausibility and adjusted where implausible. Finally, all indicators were standardized prior to analysis to eliminate the influence of differing measurement units.

Table 1. Evaluation indicator system of the water-energy-food (WEF) system.

A complete list of indicators, together with their definitions, units, and corresponding data sources, is provided in Supplementary Table S1, and the processed dataset is available from the corresponding author upon reasonable request.

2.3 Methods

2.3.1 Construction of W-E-F evaluation index system

To comprehensively assess the development level of the Water–Energy–Food (WEF) system in the Xiangjiang River Basin, a multi-dimensional evaluation index system was constructed based on the principles of system integrity, scientific validity, and data availability. Considering the interdependent nature of water, energy, and food subsystems, the index system was designed to capture both resource consumption and supply capacity, as well as efficiency and sustainability dimensions (Table 1).

Specifically, the WEF evaluation framework consists of three subsystems: water resources, energy, and food (grain). The water subsystem focuses on the balance between water supply and demand, incorporating indicators such as agricultural, industrial, and domestic water consumption, ecological water use, and water-use efficiency. These indicators reflect both the pressure on water resources and the capacity for sustainable water utilization (Gonzalez-Rosell et al., 2023).

The energy subsystem captures the scale, structure, and efficiency of energy consumption and supply. It includes indicators such as total electricity consumption, gas and liquefied petroleum gas supply, standard coal conversion coefficients, and total energy consumption. In addition, economic output (GDP) and energy intensity are incorporated to reflect the relationship between energy use and economic development, thereby highlighting both energy demand and efficiency characteristics (Wang et al., 2018).

The food and agricultural production subsystem (hereafter referred to as the food subsystem) evaluates both regional food-security outcomes and the broader agricultural resource and production base that supports them. It incorporates demand-adjusted food availability, production efficiency, agricultural land endowment, mechanization capacity, and diversified agricultural outputs. Per capita grain production, grain production efficiency, and the grain self-sufficiency rate reflect food availability and productivity; crop sown area, cultivated land area, mechanized harvesting area, and agricultural machinery power represent the agricultural production base; and grain, oil-crop, cotton, and meat outputs capture the scale and diversity of regional agricultural production. Cotton output is included not as a direct food item, but as an indicator of agricultural production structure and land-based production capacity. Therefore, the subsystem is termed the “food and agricultural production subsystem” rather than a narrowly defined grain-production subsystem (Shannak et al., 2018).

All indicators were classified as either positive indicators (where higher values indicate better system performance) or negative indicators (where higher values indicate greater resource pressure or inefficiency). This classification ensures consistency in the subsequent normalization and aggregation processes.

It should be emphasized that the water, energy, and food categories in Table 1 represent subsystem boundaries rather than homogeneous one-directional dimensions. Indicator direction was assigned according to the functional effect of each variable on sustainable subsystem performance. Resource consumption, demand pressure, and inefficient resource use were treated as negative attributes, whereas supply capacity, ecological maintenance, production security, and resource-use efficiency were treated as positive attributes. Therefore, the coexistence of positive and negative indicators within a subsystem reflects the balance between system pressure and system capacity rather than a contradiction in classification.

Aggregate, per capita, and intensity indicators were retained because they measure complementary aspects of system performance. Aggregate indicators describe absolute supply or consumption scale, whereas per capita and intensity indicators reflect demand pressure, resource availability, or efficiency after controlling for population, economic output, or land area. All indicators were converted into dimensionless values within the interval [ 0 , 1 ] before entropy calculation. Entropy weights were calculated separately within each subsystem, and the resulting water, energy, and food subsystem scores were assigned equal weights in the subsequent composite evaluation and coupling coordination analysis.

2.3.2 TOPSIS entropy weight model

To evaluate the comprehensive development level of the Water–Energy–Food (WEF) system, this study adopts a combined entropy weight–TOPSIS method, which integrates objective weighting with multi-criteria decision analysis (Li et al., 2019). All indicators are first standardized according to their attributes (positive or negative) to eliminate dimensional differences.

The entropy weight method is then applied to determine the objective weight of each indicator. The entropy value of indicator e j is calculated as Equation 1:Where p ij represents the proportion of the i-th observation under indicator j. The information utility is defined as Equation 2:and the weight of each indicator is given by Equation 3:

e j = − k ∑ i = 1 n p ij ln ( p ij ) , k = 1 ln n (1)

d j = 1 − e j (2)

w j = d j ∑ d j (3)

Based on the obtained weights, the TOPSIS method is used to construct a weighted normalized matrix. The positive ideal solution and negative ideal solution are defined, and the Euclidean distances are calculated as Equation 4:

D i + = ∑ j = 1 m w j ( x ij − x j + ) 2 , D i − = ∑ j = 1 m w j ( x ij − x j − ) 2 (4)

The relative closeness coefficient is then computed as Equation 5:Where C i ∈ [ 0 , 1 ], and a higher value indicates a higher level of subsystem or composite WEF system development.

C i = D i − D i + + D i − (5)

2.3.3 Spatial autocorrelation

To examine the spatial dependence of the WEF system, both global and local spatial autocorrelation analyses are conducted based on Moran’s I statistics (Zheng et al., 2022). Global Moran’s I is used to assess the overall spatial autocorrelation pattern and is calculated as Equation 6:Where n is the number of spatial units, x i and x j denote observed values, x ¯ is the mean, w ij is the spatial weight matrix, and S 0 = ∑ i ∑ j w ij. The value of Moran’s I ranges from −1 to 1, indicating spatial clustering, dispersion, or randomness. Statistical significance is evaluated using Z-scores and p-values. To further identify local spatial patterns, Local Moran’s I (LISA) is employed in Equation 7:

I = n S 0 · ∑ i ∑ j w ij ( x i − x ¯ ) ( x j − x ¯ ) ∑ i ( x i − x ¯ ) 2 (6)

I i = ( x i − x ¯ ) ∑ j w ij ( x j − x ¯ ) (7)

Based on the sign and significance of I i, spatial units are classified into four types: high–high (HH), low–low (LL), high–low (HL), and low–high (LH), representing different forms of spatial clustering and spatial heterogeneity. A spatial weight matrix based on geographical adjacency is constructed and row-standardized to ensure comparability across regions.

2.3.4 Coupling coordination degree model

To quantitatively evaluate the interaction and coordinated development among the water, energy, and food (WEF) subsystems, the coupling coordination degree model (CCDM) was employed (Gao and Xu, 2024). This model integrates the concepts of system coupling and coordinated development, enabling a comprehensive assessment of both the strength of inter-system interactions and the overall level of balanced development.

Let W, E, and F denote the normalized indices of the water, energy, and food subsystems, respectively. The coupling degree C, which reflects the intensity of interaction among the three subsystems, is defined as Equation 8:Where C ∈ [ 0 , 1 ]. A higher value of C indicates stronger interaction among subsystems, whereas a lower value suggests weaker coupling.

C = [ W × E × F ( ( W + E + F ) / 3 ) 3 ] 1 / 3 (8)

To capture the overall development level of the WEF system, a composite evaluation index T is constructed as Equation 9:Where α, β, and γ represent the weights assigned to the water, energy, and food subsystems, respectively. Considering the equal importance of each subsystem, this study sets α = β = γ = ( 1 / 3 ).

T = αW + βE + γF (9)

The coupling coordination degree D, which reflects both interaction intensity and coordinated development, is calculated as Equation 10:Where D ∈ [ 0 , 1 ]. A higher value of D indicates a higher level of coordinated development among the subsystems. Based on the calculated values of C and D, the coupling and coordination levels are further classified into different categories to reflect the stages of system development (see Table 2).

D = C × T (10)

Table 2. Classification criteria of coupling degree and coupling coordination degree.

Two related but distinct indices are used throughout this study and should not be conflated: the coupling degree (C), which measures only the strength of interaction among the water, energy, and food subsystems, and the coupling coordination degree (D), which captures both this interaction and the level of balanced, coordinated development among them. It should be noted that the CCD model evaluates the degree of balanced and synchronized development among subsystems rather than the explicit physical flows between them, thus providing a system-level approximation of nexus coordination.

2.3.5 XGBoost + SHAP

To identify the nonlinear driving mechanisms of the Water–Energy–Food (WEF) system and its coupling coordination degree, this study employs the Extreme Gradient Boosting (XGBoost) model combined with the SHapley Additive exPlanations (SHAP) framework (Hou et al., 2026). XGBoost is a tree-based ensemble learning algorithm that constructs an additive predictive model by iteratively fitting regression trees, which enables it to capture complex nonlinear relationships and interactions among variables,which is calculated as Equation 11:Where y ̂ i is the predicted value, K is the number of trees, and f k represents each individual regression tree. The model is trained by minimizing a regularized objective function that balances prediction accuracy and model complexity at Equations 12 and 13:Where l ( y i , y ̂ i ) denotes the loss function, T is the number of leaf nodes, w j is the weight of each leaf, and γ are regularization parameters. This structure enhances the model’s generalization ability and prevents overfitting.

y ̂ i = ∑ k = 1 K f k ( x i ) (11)

L = ∑ i = 1 n l ( y i , y ̂ i ) + ∑ k = 1 K Ω ( f k ) (12)

Ω ( f k ) = γT + 1 2 λ ∑ j = 1 T w j 2 (13)

To further interpret the contribution of each influencing factor, the SHAP method is applied based on Shapley values derived from cooperative game theory, which provides a theoretically consistent way to attribute model predictions to individual features in Equation 14:Where F denotes the set of all features and S represents any subset excluding feature j. The model output can be decomposed into the sum of individual feature contributions as Equation 15:Where ϕ 0 is the expected value of the model output and ϕ j represents the marginal contribution of each feature. Positive SHAP values indicate a positive effect on the dependent variable, whereas negative values indicate an inhibitory effect.

ϕ j = ∑ S ⊆ F ∖ { j } ∣ S ∣ ! ( ∣ F ∣ − ∣ S ∣ − 1 ) ! ∣ F ∣ ! [ f ( S ∪ { j } ) − f ( S ) ] (14)

f ( x ) = ϕ 0 + ∑ j = 1 M ϕ j (15)

3 Results

3.1 Spatial–temporal evolution patterns of the water–energy–food (WEF) system

Understanding the spatiotemporal evolution of the Water–Energy–Food (WEF) system is essential for identifying changes in resource-system performance and disparities in development trajectories among cities. Figure 2 illustrates the temporal evolution of the water, energy, food, and composite WEF indices across nine cities in the Xiangjiang River Basin from 2002 to 2022. Overall, the three subsystems exhibited markedly different trajectories. The water subsystem remained relatively stable before 2019 but became increasingly differentiated thereafter; the energy subsystem generally maintained high values while showing an overall declining tendency; and the food subsystem remained stable in most cities, except for the substantial late-stage increase observed in Loudi. These heterogeneous changes resulted in a moderate improvement in the composite WEF index accompanied by widening intercity disparities.

The water subsystem followed a pattern of initial adjustment, relative stability, and late-stage differentiation. Loudi’s water index declined from approximately 0.53 in 2002 to around 0.32 in 2008, whereas most other cities remained within the range of 0.24–0.38. Between 2009 and 2018, the indices of most cities fluctuated moderately around 0.27–0.35, indicating relatively limited intercity differences. After 2019, however, several cities experienced substantial increases. Changsha showed the most pronounced growth, with its water index rising from approximately 0.44 in 2019 to more than 0.80 in 2022. Shaoyang and Yueyang also increased markedly, reaching approximately 0.60 and 0.54, respectively. Consequently, the water subsystem shifted from relative convergence during the middle stage to pronounced spatial differentiation in the final stage.

The energy subsystem maintained comparatively high index values during most of the study period, although a gradual downward trend emerged after the mid-2010s. Most cities recorded values between 0.85 and 1.00 during the early and middle stages, with Shaoyang, Loudi, Chenzhou, Zhuzhou, and Yongzhou generally occupying the upper range. By 2022, however, the indices of several cities had declined to approximately 0.82–0.91. Yueyang and Xiangtan showed relatively large short-term fluctuations, while Changsha exhibited the most persistent decline, falling from approximately 0.64 in 2002 to around 0.31 in 2022. In contrast, Hengyang increased from approximately 0.75 in the early stage to nearly 0.90 during the middle period before declining moderately. The energy subsystem therefore remained relatively strong overall but displayed increasing divergence, particularly between Changsha and the other cities.

The food subsystem was comparatively stable and remained at a relatively low level in most cities. Between 2002 and 2018, the majority of cities recorded values below 0.15. Yueyang exhibited a gradual increase from approximately 0.09 to 0.20, whereas Hengyang, Shaoyang, Yongzhou, and Chenzhou generally remained between 0.07 and 0.12. Loudi showed a distinctly different trajectory after 2018, with its food index increasing from below 0.10 to approximately 0.54 in 2019 and reaching about 0.77 during 2020–2022. This sharp increase transformed Loudi into the dominant high-value city in the food subsystem and substantially enlarged intercity disparities during the final stage.

The composite WEF index increased moderately in most cities between 2002 and 2010 and remained relatively concentrated within the range of approximately 0.48–0.59 from 2010 to 2018. After 2019, the trajectories diverged considerably. Loudi recorded the strongest increase, reaching approximately 0.79 in 2021–2022, largely reflecting the rapid improvement in its food subsystem. Yueyang also increased steadily to approximately 0.67 in 2022, supported by simultaneous gains in its water and food indices. Shaoyang and Yongzhou maintained relatively high composite values, whereas Xiangtan remained at the lower end and declined to approximately 0.46. Although Changsha experienced a sharp improvement in its water index, its composite WEF index increased only moderately because of the continued decline in its energy subsystem. Overall, the basin’s WEF system evolved from relatively stable and convergent conditions before 2019 toward stronger subsystem heterogeneity and increasing intercity differentiation.

Figure 3 further presents the spatial–temporal evolution of the water and energy subsystems across the nine cities of the Xiangjiang River Basin in 2002, 2007, 2012, 2017, and 2022. The two subsystems exhibit clearly different spatial patterns. The water subsystem shifted from a relatively localized high-value pattern to a more dispersed and differentiated distribution, whereas the energy subsystem maintained a broadly high-level peripheral pattern but showed a gradual decline and an increasingly evident low-value concentration around Changsha.

For the water subsystem (Figure 3a), the spatial distribution changed considerably over the study period. In 2002, Loudi formed the most prominent high-value area, while most other cities were classified within the lower-value categories. By 2007, Loudi’s relative advantage had weakened, and Changsha emerged as a comparatively high-value city, whereas the southern cities generally remained at low levels. In 2012 and 2017, water-source levels increased moderately in several central and eastern cities, including Changsha, Zhuzhou, and Chenzhou, producing a more spatially dispersed pattern. The most substantial restructuring occurred by 2022. High-value areas expanded to Yueyang, Changsha, Loudi, Xiangtan, Shaoyang, and Chenzhou, while Hengyang, Yongzhou, and Zhuzhou remained relatively lower. Yueyang and Changsha occupied the highest category, accompanied by marked increases in Shaoyang and Chenzhou. Therefore, the water subsystem evolved from a localized high-value pattern dominated by Loudi toward a multicentric structure with high-value areas distributed across the northern, central, western, and southeastern parts of the basin. This spatial expansion was accompanied by stronger intercity differentiation rather than uniform basin-wide improvement.

The energy subsystem displayed a relatively stable but gradually weakening spatial structure (Figure 3b). In 2002, high-value areas were widely distributed across Loudi, Xiangtan, Zhuzhou, Shaoyang, Hengyang, Yongzhou, and Chenzhou, while Yueyang occupied an intermediate level and Changsha was already located in a relatively low-value category. This broad pattern remained largely unchanged in 2007 and 2012, indicating persistent spatial continuity in the energy subsystem. Nevertheless, a gradual decline became evident after 2012. By 2017, the energy index had decreased in several cities, particularly Changsha and Yueyang, although most southern and western cities continued to maintain relatively high levels. In 2022, Changsha became the most prominent low-value area, while Yueyang also declined compared with the earlier periods. In contrast, Loudi, Xiangtan, Zhuzhou, Shaoyang, Hengyang, Yongzhou, and Chenzhou remained within relatively high-value categories, despite a general reduction in their index levels.

Figure 4 presents the spatial–temporal evolution of the food subsystem and the composite WEF system across the Xiangjiang River Basin in 2002, 2007, 2012, 2017, and 2022. Compared with the relatively stable spatial pattern observed during the early and middle periods, both indices exhibited stronger intercity differentiation by 2022, particularly because of the rapid increase in Loudi.

The food subsystem showed limited spatial variation from 2002 to 2017 (Figure 4a). During this period, most cities remained within the two lowest index categories, indicating generally low and relatively stable food-system performance. Yueyang consistently recorded comparatively high values and formed a persistent northern high-value area, while Yongzhou and Hengyang also maintained moderately higher levels in several representative years. By contrast, Changsha, Xiangtan, Zhuzhou, Loudi, and Chenzhou generally remained in the lower-value categories. Although small adjustments occurred between 2007 and 2017, the overall spatial hierarchy changed only slightly, demonstrating considerable temporal persistence.

A pronounced restructuring emerged in 2022. Loudi shifted abruptly from a low-value city to the highest food-index category, becoming the dominant high-value center of the basin. Xiangtan also recorded a visible increase, while Yueyang continued to maintain a comparatively high level. Most southern and eastern cities, including Yongzhou and Chenzhou, remained within relatively low categories. The food subsystem therefore evolved from a weakly differentiated and relatively stable pattern into a more polarized spatial structure dominated by the exceptional increase in Loudi.

The composite WEF index displayed a more heterogeneous and dynamic spatial pattern (Figure 4b). In 2002, relatively high values were concentrated mainly in Changsha, Xiangtan, Zhuzhou, and Chenzhou, whereas Shaoyang and several western or southern cities occupied lower categories. The spatial pattern remained broadly similar in 2007, although differences among the central and southeastern cities became more evident. By 2012, Loudi’s composite index had increased, while Changsha, Xiangtan, Zhuzhou, and Chenzhou continued to constitute the principal medium- and high-value areas. In 2017, Loudi strengthened its position further and joined Changsha and Chenzhou as one of the relatively high-performing cities, while Shaoyang, Yueyang, and Yongzhou remained at lower levels.

By 2022, the spatial hierarchy of the composite WEF system had been substantially reorganized. Loudi entered the highest index category and became the most prominent high-value center, primarily corresponding to its sharp increase in the food subsystem and improvement in the water subsystem. Yueyang and Shaoyang also moved into relatively high categories, producing a more dispersed distribution of high values across the northern and western parts of the basin. In contrast, Changsha declined to a comparatively low composite level despite its substantial improvement in the water subsystem, reflecting the constraining effect of its weak energy performance. Yongzhou also remained at a relatively low level, whereas Xiangtan, Zhuzhou, Hengyang, and Chenzhou occupied intermediate positions.

Overall, the composite WEF system evolved from a relatively stable spatial pattern dominated by central and southeastern cities into a more dispersed and polarized structure by 2022. The emergence of Loudi as the leading city and the contrasting decline of Changsha indicate that changes in the composite index were shaped by different combinations of subsystem performance rather than by uniform improvement across water, energy, and food systems. To further examine changes in distribution shape, dispersion, and polarization, kernel density estimation was subsequently applied to the subsystem and composite WEF indices (Figure 5).

Figure 5 illustrates the dynamic distribution of the water, energy, food, and composite WEF indices from 2002 to 2022. Overall, the four indices exhibit distinct distributional trajectories. The water and food subsystems are mainly concentrated at relatively low values but develop increasingly evident right tails and secondary peaks in the later period. The energy subsystem remains concentrated at relatively high values, although its distribution gradually shifts leftward and becomes more dispersed. The composite WEF index is primarily distributed within the middle-value range, with increasing multimodality and widening intercity disparities over time.

The water and food subsystems display broadly similar patterns of late-stage differentiation. The water index is mainly concentrated between approximately 0.25 and 0.40, with only limited movement in the main peak during most of the study period. After 2019, however, the distribution becomes wider and secondary peaks emerge above 0.50, reflecting the rapid improvement of a small number of cities. Similarly, the food index remains concentrated mainly between approximately 0.03 and 0.15, indicating relatively stable and low-level performance in most cities. In the later years, isolated peaks appear at substantially higher values, largely corresponding to the sharp increase in Loudi. These results indicate that both subsystems evolved from relatively concentrated distributions toward stronger right-skewness and polarization, rather than uniform basin-wide improvement.

The energy subsystem shows an opposite tendency. Its dominant peak is initially concentrated between approximately 0.85 and 1.00, but gradually shifts toward lower values and becomes less concentrated over time. The emergence of secondary peaks between approximately 0.40 and 0.85 indicates increasing differentiation between the majority of cities that maintained relatively high levels and a small number of cities that experienced substantial declines. The energy subsystem therefore exhibits a general downward adjustment accompanied by growing dispersion.

The composite WEF index is mainly concentrated between approximately 0.45 and 0.60. During the early and middle periods, the distribution is relatively compact, although several local peaks indicate persistent intercity heterogeneity. In the later stage, the distribution widens and develops a more pronounced right tail, with secondary peaks extending toward approximately 0.65–0.80. This suggests that the integrated WEF performance of some cities improved substantially, while others remained within the middle- or lower-value ranges. Overall, the kernel density results reveal no consistent convergence across the basin; instead, improvements in the composite WEF system were accompanied by increasing subsystem heterogeneity and intercity polarization.

3.2 Spatial autocorrelation analysis of the WEF system

The Global Moran’s I results indicate that the water subsystem generally exhibited weak and statistically insignificant spatial dependence from 2002 to 2022 (Table 3). Moran’s I was negative in most years, ranging from −0.3479 in 2014 to 0.3963 in 2020. During 2002–2019, the predominantly negative values suggest a weak tendency toward spatial dispersion, meaning that cities with relatively high and low water-source levels were more often located adjacent to dissimilar cities. However, none of these values was statistically significant. A notable exception occurred in 2020, when Moran’s I increased sharply to 0.3963 and became statistically significant at the 1% level (p = 0.0056), indicating temporary positive spatial clustering. This clustering did not persist, as Moran’s I returned to −0.0982 in 2021 and 0.0040 in 2022, both statistically insignificant. Overall, the water subsystem did not form a stable global spatial cluster, although a short-lived concentration emerged in 2020.

Table 3. Xiangjiang river basin water sources and energy sources level global Moran’s I.

The energy subsystem similarly displayed weak global spatial autocorrelation over most of the study period. Moran’s I was negative in nearly all years, ranging from −0.3007 in 2007 to 0.0732 in 2020, indicating a general tendency toward spatial dispersion rather than clustering. The only clearly positive values occurred in 2012, 2019, and 2020, but none reached the conventional 5% significance threshold. In 2020, Moran’s I rose to 0.0732 with a p-value of 0.0594, suggesting weak positive spatial dependence at the 10% significance level. Nevertheless, the index returned to negative values in 2021 and 2022. These results indicate that the energy subsystem also lacked a persistent basin-wide clustering pattern, despite a temporary weakening of spatial dispersion around 2019–2020.

Taken together, the two subsystems were characterized mainly by unstable and weak spatial dependence. The water subsystem exhibited one significant episode of positive clustering in 2020, whereas the energy subsystem showed only marginal evidence of clustering in the same year. The limited significance of the Global Moran’s I statistics may partly reflect the small number of spatial units in the basin, which reduces the statistical power of the test. It may also indicate that the spatial differences identified in the maps are primarily localized or city-specific rather than organized into stable basin-wide clusters. Therefore, local spatial association analysis is needed to determine whether particular high–high, low–low, or spatial-outlier configurations exist despite the generally insignificant global results.

The results for the food subsystem and the composite WEF system further confirm that global spatial dependence was generally weak and statistically insignificant during 2002–2022 (Table 4).

Table 4. Xiangjiang river basin food sources and composite water-energy-food system level global Moran’s I.

For the food subsystem, Global Moran’s I remained close to zero during most of the study period. From 2002 to 2017, the values fluctuated narrowly between −0.0104 and 0.0532, indicating an essentially random spatial distribution with no stable clustering or dispersion pattern. A modest increase appeared from 2018 to 2019, when Moran’s I rose to 0.0610 and 0.0879, respectively, but the corresponding p-values remained well above 0.05. After 2020, Moran’s I turned negative, reaching −0.1237 in 2020, −0.1291 in 2021, and −0.1299 in 2022. However, these values were very close to the expected value of −0.125 and remained statistically insignificant. Overall, the food subsystem exhibited neither persistent positive clustering nor meaningful spatial dispersion at the basin scale.

The composite WEF system displayed relatively stronger positive spatial autocorrelation in the early years, although the results were still mostly insignificant. Moran’s I reached 0.2418 in 2002, with a p-value of 0.0633, indicating weak evidence of positive spatial clustering at the 10% significance level. From 2003 to 2019, the index remained positive but generally declined and fluctuated between 0.0456 and 0.1794. This suggests a persistent but weak tendency for cities with similar composite WEF levels to be spatially adjacent, without sufficient statistical evidence to confirm stable basin-wide clustering. A marked change occurred in 2020 and 2021, when Moran’s I became negative at −0.1261 and −0.1609, respectively. Nevertheless, both values were statistically insignificant and were followed by a return toward zero in 2022, when Moran’s I was −0.0104.

Taken together, the composite WEF system did not exhibit the previously assumed transition from dispersion to increasing clustering. Instead, it showed a weak positive clustering tendency during much of 2002–2019, followed by a temporary shift toward spatial dispersion in 2020-2021. The food subsystem remained predominantly random throughout the study period. These findings suggest that the spatial differences observed in the thematic maps were mainly localized or driven by individual cities rather than forming stable global clusters. Therefore, the absence of significant Global Moran’s I values does not imply a complete lack of spatial heterogeneity, but rather indicates that such heterogeneity was not consistently organized at the basin-wide scale.

To further examine localized spatial relationships that may not be reflected by Global Moran’s I, Moran scatterplots were used to classify cities into high–high (HH), low–low (LL), high–low (HL), and low–high (LH) quadrants for 2002, 2012, and 2022 (Figure 6). Overall, the cities were distributed across all four quadrants, and many observations were located close to the origin. This indicates weak and unstable local spatial associations rather than persistent basin-wide clustering. Moreover, because Figure 6 does not report the significance of city-specific Local Moran’s I statistics, the quadrant positions should be interpreted as descriptive spatial associations rather than statistically confirmed LISA clusters.

For the water subsystem (Figure 6a), local spatial relationships changed considerably across the three representative years. In 2002, Xiangtan, Yueyang, and Hengyang were mainly located in the LH quadrant, whereas Changsha and Shaoyang displayed HH characteristics. Loudi formed a marked HL spatial outlier, reflecting its relatively high water index compared with the lower values of surrounding cities. By 2012, Yueyang remained a prominent LH city, while Changsha shifted to the HL quadrant. In 2022, Yueyang continued to exhibit an LH relationship, whereas Changsha and Shaoyang were positioned in the HL quadrant. Several other cities remained close to the origin. These changes suggest that the water subsystem did not develop a stable local clustering structure; instead, it was characterized by shifting spatial outliers and increasingly heterogeneous city–neighbor relationships.

The energy subsystem displayed comparatively persistent local contrasts (Figure 6b). Changsha was located in the LH quadrant in all three representative years, indicating that its relatively low energy index was surrounded by cities with higher average values. Yueyang, by contrast, generally exhibited an HL relationship, particularly in 2002, 2012, and 2022, because its energy performance differed from that of its neighboring cities. Several southern and central cities, including Hengyang and Chenzhou, tended to occupy the HH quadrant in the later years, although most observations remained relatively close to the horizontal axis. The results therefore indicate a persistent low-value spatial outlier centered on Changsha, rather than the high-value growth-pole pattern previously assumed.

For the food subsystem (Figure 6c), most cities remained relatively close to the origin during 2002 and 2012, consistent with the weak Global Moran’s I values. Shaoyang and Yongzhou generally occupied the HH quadrant, while Changsha and several central cities tended to fall within the LL quadrant. However, the quadrant configuration changed markedly by 2022 following the sharp increase in Loudi’s food index. Loudi became a pronounced HL outlier, indicating a high local value surrounded by comparatively lower-value neighbors. Meanwhile, several other cities shifted toward the LH or LL quadrants. Thus, the food subsystem changed from a weakly differentiated local pattern to one dominated by an isolated high-value city rather than a broader high-value cluster.

The composite WEF system also exhibited changing and mixed local spatial associations (Figure 6d). In 2002, Shaoyang and Yongzhou were located in the HH quadrant, while Changsha, Zhuzhou, and Xiangtan showed LL characteristics. In 2012, the overall distribution remained dispersed, with several southern cities occupying relatively high-value quadrants and central cities remaining in LL or LH positions. By 2022, Loudi became the most evident HL outlier, corresponding to its rapid increase in the composite WEF index, whereas Shaoyang retained an HH relationship and Zhuzhou remained in the LL quadrant. Yueyang and several southern cities were mainly located in the HL quadrant. Overall, the composite system did not evolve toward a stable “core–periphery” structure. Instead, it was characterized by the coexistence of weak clusters, spatial outliers, and frequent changes in quadrant membership, consistent with the generally insignificant global spatial autocorrelation results.

3.3 Spatial–temporal evolution of the coupling coordination degree of the WEF system

Figures 7 and 8 present the temporal evolution of the WEF coupling degree (C) and coupling coordination degree (D) across the upstream, midstream, and downstream regions of the Xiangjiang River Basin from 2002 to 2022. Overall, the coupling degree remained relatively high, with most cities located in the running-in stage, whereas the coordination degree showed greater differentiation and more evident stage transitions. The results indicate that strong interactions among the water, energy, and food subsystems did not necessarily correspond to equally high coordinated development.

As shown in Figure 7, the upstream cities remained within the running-in stage throughout the study period. Chenzhou’s coupling degree fluctuated moderately between approximately 0.57 and 0.63, displaying a gradual upward trend after 2010. Yongzhou consistently recorded higher values, increasing from about 0.64 in 2002 to above 0.70 after 2007, despite several short-term fluctuations. It reached a temporary peak of approximately 0.76 in 2021 before declining slightly in 2022. These trajectories indicate relatively stable subsystem interactions in the upstream region, with Yongzhou exhibiting stronger coupling than Chenzhou.

The midstream region displayed the greatest internal differentiation. Hengyang maintained a comparatively high coupling degree of approximately 0.71–0.78, although its value declined slightly after its early peak. Shaoyang followed a gradual upward trajectory, increasing from about 0.64 to approximately 0.71 while remaining in the running-in stage. Loudi exhibited the most pronounced structural change. Its coupling degree remained near the lower boundary of the running-in stage during most of 2002–2019 and briefly entered the antagonistic stage in the early period. After 2019, however, it increased sharply to above 0.90 and entered the high-level coupling stage. This abrupt increase corresponds to the substantial late-stage changes in Loudi’s food and composite WEF indices.

In the downstream region, two distinct groups were evident. Changsha and Yueyang maintained relatively high coupling levels and temporarily entered the high-level coupling stage. Changsha increased from approximately 0.70 to a peak of about 0.85 in the mid-2010s, followed by a marked decline to around 0.74 in 2022. Yueyang rose relatively steadily from approximately 0.70 to above 0.80 in the later period. By contrast, Zhuzhou and Xiangtan remained in the lower portion of the running-in stage. Xiangtan experienced an early increase to approximately 0.65 but subsequently declined, while Zhuzhou fluctuated narrowly around 0.55–0.59. Therefore, the downstream region did not exhibit uniformly superior coupling performance, but rather a polarized structure dominated by Changsha and Yueyang.

Figure 8 shows that the coupling coordination degree was lower than the coupling degree but generally improved over time. In the upstream region, Chenzhou fluctuated mainly around the boundary between the verge of maladjustment and barely coordinated stages, increasing from approximately 0.49 to around 0.53. Yongzhou performed better, rising from about 0.52 to approximately 0.57 and briefly exceeding 0.60 in 2021. Although Yongzhou subsequently declined, its trajectory indicates a gradual transition toward primary coordination, whereas Chenzhou remained at a relatively low coordination level.

The midstream region again exhibited substantial heterogeneity. Hengyang remained comparatively stable at approximately 0.52–0.57, corresponding mainly to the barely coordinated stage. Shaoyang increased gradually from around 0.54 to above 0.60 by 2022 and entered the primary coordination stage. Loudi remained near 0.48–0.50 for most of the study period, indicating persistent maladjustment or only marginal coordination. After 2019, however, its coordination degree rose sharply to approximately 0.74 in 2020 and nearly 0.80 by 2022, representing the most significant improvement among all cities and a transition toward intermediate coordination.

In the downstream region, Yueyang exhibited the clearest and most sustained improvement, increasing from approximately 0.52 in 2002 to about 0.67 in 2022 and entering the primary coordination stage. Changsha also improved moderately, rising from around 0.49 to approximately 0.55, although its trajectory was interrupted by several fluctuations. Zhuzhou and Xiangtan remained comparatively low. Zhuzhou fluctuated close to 0.50 and showed a slight downward tendency in the later years, while Xiangtan declined from an early peak above 0.50 to approximately 0.46 by 2022. Thus, despite the relatively advanced economic development of the downstream region, coordinated WEF performance remained uneven.

Overall, most cities maintained medium-to-high coupling degrees, but their coordination levels were considerably more differentiated. Loudi experienced the strongest late-stage transition in both coupling and coordination, while Yueyang demonstrated the most sustained improvement. Yongzhou and Shaoyang also showed gradual progress, whereas Chenzhou, Zhuzhou, and Xiangtan remained at relatively low coordination stages. Changsha’s declining coupling degree after the mid-2010s and only moderate coordination improvement further demonstrate that intense subsystem interactions alone do not guarantee balanced WEF development.

3.4 Driving factors of the composite water-energy-food system level and coupling coordination degree of the WEF system

The identification of driving factors is essential for understanding the mechanisms underlying the spatial–temporal evolution of the WEF system and its coupling coordination degree. Based on existing literature on resource nexus systems, regional sustainability, and urban–rural development, this study selects a set of socio-economic, environmental, and structural variables that are theoretically linked to water, energy, and food interactions.

Specifically, the influencing factors are categorized into five dimensions to comprehensively capture the drivers of the WEF system and its coupling coordination degree (Table 5).

Table 5. Influencing factors of the WEF system and coupling coordination degree.

First, economic development variables (X1–X3) are included to represent the scale and fiscal capacity of regional economies. Higher economic development generally enhances infrastructure investment and resource allocation efficiency, thereby promoting system coordination. Second, industrial structure and openness variables (X4, X5) reflect the transformation of economic structure and the degree of external integration. A higher share of the tertiary industry and greater openness are expected to improve resource efficiency and reduce pressure on traditional resource-intensive sectors. Third, urbanization and demographic factors (X6–X8) are incorporated to capture population concentration and spatial expansion. While urbanization may improve technological efficiency and service provision, excessive population density and scale can increase pressure on water, energy, and food systems. Fourth, agricultural production capacity variables (X9–X11) are used to characterize the supply-side capacity of the food subsystem. Higher levels of mechanization and cultivated land input are expected to enhance agricultural productivity and contribute positively to system stability. Fifth, resource, environmental, and efficiency-related variables (X12–X16) reflect both natural constraints and technological progress. Per capita water use and energy intensity capture resource consumption pressure, while green coverage and science and technology investment represent ecological and innovation capacity that may facilitate coordinated development.

Figure 9 presents the SHAP-based importance rankings of the influencing factors for the coupling coordination degree (Y1) and the composite WEF system level (Y2). The mean absolute SHAP value represents the average contribution of each variable to model prediction, with larger values indicating stronger explanatory power. It should be noted that feature importance reflects the magnitude of influence rather than the direction of effects.

For the coupling coordination degree (Y1), per capita water use is identified as the most influential factor, with the highest mean absolute SHAP value (approximately 0.28), indicating that water-use efficiency and resource utilization patterns play a critical role in regulating interactions among water, energy, and food subsystems. The share of tertiary industry and urbanization rate rank second and third, respectively, suggesting that industrial transformation and urban development processes substantially influence the coordination capacity of the WEF system. Agricultural mechanization level also demonstrates considerable importance, highlighting the role of agricultural modernization in improving subsystem integration. Other socioeconomic and resource-related variables, including resident population, GDP per capita, energy intensity, and GDP, show moderate contributions, whereas crop sown area, agricultural machinery power, fiscal expenditure, and technological indicators exhibit relatively weaker explanatory power. These results indicate that coupling coordination is primarily driven by resource-use efficiency, industrial restructuring, and urbanization dynamics rather than by individual agricultural production factors alone.

For the composite WEF system level (Y2), the importance structure differs substantially from that of Y1. Crop sown area emerges as the dominant factor, with the highest mean absolute SHAP value (approximately 0.21), demonstrating the fundamental role of agricultural resource endowment in determining overall WEF system performance. Per capita water use and resident population are the second and third most important factors, respectively, indicating that resource consumption intensity and demographic pressure are key determinants of the integrated WEF level. Agricultural machinery power also exhibits strong explanatory power, reflecting the contribution of agricultural productivity and technological conditions. In addition, the share of tertiary industry, fiscal expenditure, GDP, and agricultural mechanization level provide moderate contributions, suggesting that economic development and public investment further shape the overall resource-system capacity. Compared with Y1, the importance distribution of Y2 is more balanced, involving a broader combination of natural resource conditions, socioeconomic scale, and production capacity.

A comparison between the two models reveals distinct driving mechanisms. Although per capita water use is among the most important factors for both outcomes, its role differs between system coordination and overall system level. For Y1, water-use characteristics are closely associated with the balance and interaction among subsystems, together with industrial restructuring and urbanization. For Y2, agricultural resource capacity, particularly crop production conditions and population demand, plays a more dominant role in determining the overall WEF development level. These differences suggest that improving WEF coordination requires not only enhancing resource productivity but also optimizing industrial and urban development patterns, whereas improving the overall WEF system level depends more strongly on strengthening agricultural capacity, resource management, and socioeconomic support.

The SHAP summary plots reveal marked differences in both the direction and dispersion of the drivers affecting the coupling coordination degree (Y1) and the composite WEF system level (Y2) (Figure 10). Each point represents an observation, the horizontal position indicates its contribution to the predicted outcome, and the color reflects the feature value, from low (blue) to high (red). Positive SHAP values increase the predicted index, whereas negative values reduce it.

For the coupling coordination degree (Y1; Figure 10a), per capita water use exhibits the widest SHAP range, although its effect is strongly nonlinear. Most observations are concentrated near zero, while several low-value observations generate exceptionally large positive contributions. Resident population and crop sown area show clearer positive relationships: high feature values are generally located on the positive side of the SHAP axis, whereas low values are concentrated near zero or on the negative side. Agricultural mechanization level and agricultural machinery power also contribute positively, with several high-value observations producing relatively large positive SHAP values. The share of tertiary industry displays a similar but more heterogeneous pattern, including a small number of strong positive effects. By contrast, most remaining variables, including fiscal expenditure, GDP per capita, green coverage, openness, science expenditure, technology expenditure, energy intensity, and population density, are tightly clustered around zero, indicating relatively limited average marginal effects.

For the composite WEF system level (Y2; Figure 10b), crop sown area displays the clearest positive relationship. High values are predominantly associated with positive SHAP contributions, while low values generate negative effects. Agricultural machinery power also shows a strong positive pattern, indicating that greater agricultural production capacity is associated with a higher composite WEF level. Resident population generally contributes positively, although its effects are more dispersed. Per capita water use again demonstrates a nonlinear relationship: most high-value observations remain close to zero or exert slightly negative effects, whereas a few low-value observations produce substantial positive contributions. The share of tertiary industry, degree of openness, agricultural mechanization level, and fiscal expenditure show mainly positive effects at higher values. In contrast, higher population density and urbanization rates tend to be associated with negative SHAP values, suggesting that intensified population concentration and urban expansion may place pressure on the integrated WEF system. Energy intensity also shows a predominantly weak or negative contribution, while GDP per capita, science expenditure, green coverage, and technology expenditure exert comparatively small effects.

Overall, the coupling coordination degree is influenced by several strongly nonlinear and observation-specific effects, particularly those associated with per capita water use and agricultural production conditions. The composite WEF system level shows clearer directional relationships, with crop sown area, agricultural machinery power, and resident population generally exerting positive effects, while population density, urbanization, and energy intensity tend to constrain system performance. These results indicate that agricultural resource endowment and production capacity remain central to both outcomes, but socioeconomic concentration and resource-use pressure affect the composite WEF level more clearly.

4 Discussion

The findings of this study reveal that the WEF system in the Xiangjiang River Basin is characterized by a process of gradual improvement accompanied by persistent spatial heterogeneity and structural imbalance. Although the composite system level has steadily increased over time, the coupling coordination degree remains at a relatively low level, with most cities still confined to the stage of moderate maladjustment. This divergence between system development and coordination suggests that improvements in individual subsystems do not automatically translate into integrated and synergistic outcomes, highlighting the importance of examining both system performance and coordination simultaneously.

The results further indicate that the underlying mechanisms of system coordination and system level differ substantially. The SHAP analysis shows that the coupling coordination degree is primarily driven by agricultural production capacity and resource-use efficiency, particularly agricultural mechanization, machinery power, and energy intensity. This suggests that coordination within the WEF system depends largely on the alignment between production processes and resource consumption efficiency. In contrast, the composite WEF system level is influenced by a broader set of factors, including technological investment and fiscal support, indicating that overall system performance is shaped by both production capacity and institutional or policy-related inputs. This distinction reflects the fundamental difference between “system growth” and “system coordination”: while growth can be achieved through expansion and investment, coordination requires structural optimization and efficiency improvements across subsystems.

From a spatial perspective, the weak global spatial autocorrelation but evident local clustering patterns suggest that the WEF system exhibits a fragmented yet locally structured spatial organization. The emergence of high–low (HL) patterns, particularly in cities such as Changsha, indicates that leading cities function as growth poles but generate limited spillover effects to surrounding areas. This finding implies that regional development is characterized by a core-driven but weakly diffused structure, which may exacerbate spatial inequality. Meanwhile, the relatively stable and weakly differentiated food subsystem highlights the strong dependence of agricultural production on natural conditions and long-term land-use structures, further constraining spatial convergence within the WEF system.

These findings are consistent with previous studies emphasizing the importance of technological progress and efficiency improvements in resource nexus systems, while also extending the literature by distinguishing between system level and coordination mechanisms. Existing research has often focused on either resource efficiency or system performance, whereas this study demonstrates that coordination outcomes are more strongly associated with production efficiency and structural alignment rather than economic scale alone (Huysmans et al., 2015). Moreover, the identified spatial patterns support the argument that regional resource systems often evolve through uneven development processes, where leading regions advance rapidly but fail to generate sufficient spillover effects (Döring and Schnellenbach, 2006). These results both echo and qualify recent basin-scale evidence. In a closely related study of the Yellow River Basin (YRB), Zheng et al. (2025) likewise reported persistently low WEF coordination alongside strong upstream–midstream–downstream differentiation, and, using a Pressure–State–Response structural model, concluded that human and institutional “response” factors, particularly innovation, act as key enhancers of coordination in the more developed downstream reaches, whereas terrain-induced pressures constrain it upstream. Our findings for the Xiangjiang Basin partly converge with this picture: the most economically advanced downstream cities, led by Changsha, also attain the highest coordination. However, our XGBoost–SHAP decomposition refines the mechanism. We find that institutional and investment-type inputs (science-and-technology expenditure, fiscal support, economic scale) are associated primarily with the composite system level rather than with coordination, whereas coordination is governed mainly by production-side efficiency factors such as agricultural mechanization and energy intensity. Thus, where Zheng et al. emphasize institutional response as a driver of coordination, our humid, water-abundant, energy-led basin, in contrast to the arid, terrain-constrained YRB which suggests that such institutional inputs chiefly raise overall system performance, while coordination hinges on the structural alignment of production and resource use. This contrast indicates that the determinants of WEF coordination are context-dependent, and that conflating system growth with system coordination, as much of the existing literature implicitly does, can obscure these distinct mechanisms.

A further consideration concerns the alignment between our methodological design and the theoretical connotation of the WEF nexus. As originally conceptualized by Hoff (2011) and Biggs et al. (2012), the nexus is defined by the cross-sectoral trade-offs and synergies arising from the physical interdependence of water, energy, and food flows (Hoff, 2011). The coupling coordination degree (CCD) model adopted here primarily measures the degree of synchronization among subsystem development levels, rather than the magnitude or direction of resource flows between subsystems. We therefore interpret the CCD not as a direct measure of physical resource exchange, but as a system-level indicator of whether the three subsystems develop in a balanced and mutually reinforcing manner, that is, as an aggregate manifestation of the underlying nexus interactions. Importantly, this interactive dimension is partially recovered through our XGBoost–SHAP driver analysis, since the dominant determinants of coordination are themselves intrinsically cross-sectoral. For example, agricultural mechanization represents an energy input to food production, while energy intensity links energy use to resource-use efficiency. These relationships indicate that inter-subsystem dependencies, although not modeled as explicit flows, are still reflected in the drivers of coordinated development. We nonetheless acknowledge that a CCD-and-indicator framework cannot fully capture the bidirectional resource flows (e.g., water embodied in energy generation, energy consumed in water treatment and food processing) that constitute the physical core of the nexus, and that flow-based approaches such as input–output and ecological network analysis remain necessary to characterize these interlinkages directly.

The results have important policy implications for promoting sustainable and coordinated development of the WEF system. First, enhancing agricultural modernization and improving resource-use efficiency should be prioritized, as these factors play a decisive role in strengthening system coordination. Second, increasing investment in science and technology and optimizing fiscal support mechanisms can effectively improve the overall system level. Third, region-specific strategies are necessary to address spatial disparities, particularly by strengthening interregional linkages and enhancing the diffusion of technological and infrastructural benefits from core cities to peripheral areas. Finally, integrated policy frameworks that simultaneously target water, energy, and food subsystems are essential to overcome structural imbalances and promote synergistic development.

The findings contribute to the growing WEF nexus literature in three aspects. First, this study distinguishes between system performance and system coordination, revealing that the two dimensions are driven by different mechanisms. Second, the results demonstrate that agricultural modernization and resource-use efficiency are more critical for coordination than for system growth. Third, the study highlights the existence of local spatial clustering despite weak global spatial dependence, suggesting that resource nexus systems evolve through uneven and path-dependent regional development processes.

5 Conclusion

This study systematically examined the spatiotemporal evolution, spatial dependence, and driving mechanisms of the Water–Energy–Food (WEF) system in the Xiangjiang River Basin from 2002 to 2022. By integrating the entropy weight–TOPSIS method, spatial autocorrelation analysis, coupling coordination degree model, and XGBoost–SHAP framework, the study provides a comprehensive understanding of both system performance and coordination dynamics. The results reveal that the composite WEF system level exhibits a steady upward trend, primarily driven by the continuous improvement of the energy subsystem, while the water subsystem shows a fluctuating recovery pattern and the food subsystem remains relatively stable. Despite this overall improvement, significant spatial heterogeneity persists, with a clear core–periphery structure characterized by higher levels in downstream cities and lower levels in central and southwestern regions.

Second, the coupling coordination degree demonstrates a gradual transition from antagonistic stages toward the running-in stage; however, the overall coordination level remains relatively low, with most cities still experiencing moderate maladjustment. This indicates that improvements in subsystem performance do not automatically translate into coordinated development, reflecting structural imbalances within the WEF system. Spatial analysis further shows weak global spatial autocorrelation but evident local clustering patterns, suggesting that the system is characterized by fragmented spatial organization and limited spillover effects from leading cities. This highlights the importance of addressing regional disparities and enhancing interregional connectivity in resource management.

Finally, the XGBoost–SHAP analysis reveals significant differences in the driving mechanisms of system coordination and system level. The coupling coordination degree is primarily driven by agricultural production capacity and resource-use efficiency, particularly agricultural mechanization and energy intensity, indicating the critical role of production-side optimization. In contrast, the composite WEF system level is influenced by a broader combination of factors, including technological investment, fiscal support, and economic development, reflecting a more complex and diversified influence structure. These findings suggest that achieving sustainable and coordinated development of the WEF system requires not only improving individual subsystem performance but also enhancing structural alignment and efficiency across subsystems. Accordingly, policy efforts should prioritize agricultural modernization, technological innovation, and region-specific strategies to reduce spatial disparities and promote integrated resource management.

Despite the contributions of this study, several limitations should be acknowledged. First, the analysis is conducted at the prefecture-level scale, which may mask finer spatial heterogeneity within cities and limit the ability to capture intra-urban disparities in the Water–Energy–Food system. Second, although the constructed indicator system covers key dimensions of water, energy, and food subsystems, certain qualitative aspects such as institutional governance, policy intervention, and environmental regulation are not explicitly incorporated due to data constraints. Third, while the XGBoost–SHAP framework effectively captures nonlinear relationships and provides interpretability, it remains limited in revealing causal mechanisms, and therefore the identified driving factors should be interpreted as associations rather than strict causality.

Future research can be extended in several directions. First, multi-scale analyses incorporating county-level or gridded data could be conducted to reveal more detailed spatial heterogeneity within the basin. Second, integrating causal inference methods or system dynamics modeling may further strengthen the explanation of dynamic interactions among water, energy, and food subsystems. Third, scenario-based simulations could be introduced to evaluate the potential impacts of policy interventions, climate change, and land-use transitions on WEF system coordination. Finally, comparative studies across multiple river basins would help to generalize the findings and improve the transferability of the proposed analytical framework. Furthermore, the coupling coordination degree model characterizes the synchronization of subsystem development levels rather than the explicit physical resource flows—such as water-for-energy and energy-for-food transfers—that define the WEF nexus in its strict sense. Future work could integrate input–output or ecological network analysis to quantify these inter-subsystem flows directly, complementing the coordination-based perspective adopted here.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frwa.2026.1880011/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Reza Kerachian, University of Tehran, Iran

Reviewed by: Shenglin Ma, North University of China, China

Mingguang Li, Huaqiao University, China