Evolution Characteristics of Water Use Efficiency and the Impact of Its Driving Factors on the Yunnan–Guizhou Plateau in China

Abstract

Water use efficiency (WUE) of ecosystems plays a crucial role in balancing carbon storage and water consumption. The Yunnan–Guizhou Plateau, a karst landscape region with relatively fragile ecosystems in China, requires a better understanding of the evolution of WUE and the factors driving it for the region’s ecological sustainability. This study employs Theil–Sen slope estimation and Mann–Kendall significance analysis to investigate the temporal trends and spatial patterns of WUE in the study area. Additionally, a machine learning model, XGBoost, is used to establish driving relationships, and the SHAP model is applied to interpret the importance of the driving factors and their specific relationship with WUE. The results show that (1) WUE exhibits an increasing trend, with a slope of 0.002, indicating improved water absorption and utilization capacity of vegetation in the region. (2) The spatial distribution of WUE follows a “high–low–high” pattern from southwest to northeast, with 6.68% of the area showing a significant increase, 50.80% showing a weak increase, 4.60% showing a significant decrease, and 37.92% showing a weak decrease. (3) The importance of the driving factors is ranked as follows: normalized difference vegetation index (NDVI), maximum temperature (TMAX), shortwave radiation (SRAD), Palmer drought severity index (PDSI), vapor pressure deficit (VPD), and precipitation (PRE). The NDVI has a linear positive relationship with WUE; SRAD has a decreasing effect on WUE, with this effect weakening at higher values; and TMAX, PRE, the PDSI, and VPD show a non-monotonic relationship with WUE, increasing and then decreasing. The findings of this study are significant for ecological civilization construction and sustainable development in the region.

1. Introduction

Water use efficiency (WUE) in an ecosystem is the ratio of gross primary productivity (GPP) to evapotranspiration (ET), reflecting the balance between carbon storage and water consumption within the system [1]. As a key indicator for measuring ecological sustainability, WUE reflects the process of carbon–water exchange between terrestrial ecosystems and the atmosphere, and can quantify the amount of water consumed by ecosystems during carbon assimilation [2,3]. WUE plays a crucial role in the global ecosystem’s functionality, ecological services, and its feedback to climate change [4,5]. Therefore, a deeper understanding of the spatiotemporal variations in WUE and its driving factors can contribute to promoting regional sustainable development and the effective use of water resources.

Traditional methods for obtaining WUE include field measurements, data mining, and process-based model simulations. Among these, field measurements require significant human and material resources, making them less efficient and suitable only for smaller study areas [6,7]. Data mining and process models are limited by parameters and are highly dependent on input data, which may lead to biased results [8,9]. With the development of remote sensing technologies and inversion methods, it is now easier to obtain high-precision GPP and ET products. Currently, remote sensing data is one of the mainstream methods for obtaining GPP and ET on a larger scale, and using remote sensing data to estimate WUE has been widely applied in various ecological and environmental studies [10,11,12]. Examples include small-scale studies on agricultural land [13], on grasslands [14], and in forests [15]; medium-scale studies in cities such as Beijing [16], North China [17], and Xinjiang [18]; and large-scale studies across regions like India [19], the Northern Hemisphere [20], and globally [21].

For the analysis of the driving factors of WUE Xiao, et al. [22] used regression methods to analyze the impacts of climate change, vegetation growth, and CO₂ concentrations on the interannual variations in WUE. Ji et al. [23] explored the effects of industrial structure and urban environments in the Yangtze River basin on WUE using geographically weighted regression. Yang et al. [24] employed Spearman correlation to analyze the drought response and dominant ecosystem processes of WUE using four different WUE products. Cai et al. [25] used geographic detectors to study the influence of meteorological factors on WUE in the northernmost region of the Yellow River, finding that the interaction of multiple climatic factors had a stronger effect on WUE than any single factor. Jia et al. [26] explored the impact of CO₂ fertilization on WUE in the Tibetan Plateau region using a land biosphere model, concluding that CO₂ fertilization had a more significant effect on WUE in the driest years compared to the annual average. Most of these traditional methods for analyzing driving factors focus on correlation or simple regression analysis and have shown that the spatiotemporal variations in various driving factors have a substantial impact on WUE. However, the impact of each driving factor on WUE dynamics still requires further investigation. The changes in WUE cannot be explained solely by the strength of the correlation with the driving factors; the trends, including both the magnitude and the direction of change, must also be taken into account. The contribution of each driving factor to the WUE trend involves not only the relationship’s effect but also the direction and extent of the driving factor’s trend. Machine learning models such as XGBoost can efficiently establish nonlinear relationships between driving factors and WUE, providing a more reasonable representation of actual trends. Furthermore, the interpretable model SHAP, which was introduced in recent years to address the issue of machine learning’s lack of interpretability, can be used to explain the specific relationships between driving factors and WUE by leveraging the interpretability of SHAP in conjunction with XGBoost.

The southwestern region of China, particularly the Yunnan–Guizhou Plateau, is a typical karst landscape area with a relatively fragile ecosystem. In a plateau environment such as the Yunnan–Guizhou Plateau, due to its unique geological structure and complex hydrological and climatic conditions, the soil is thin and poorly permeable, and water resources are unevenly distributed, often leading to challenges such as water scarcity and ecological degradation. it is necessary to clarify the evolution characteristics of WUE in this region and the factors that drive these changes. Therefore, this study focuses on the Yunnan–Guizhou Plateau, calculating the WUE for the years 2003–2020 using the GPP and ET ratio methods. The spatiotemporal evolution of WUE in this region is explored using Theil–Sen slope estimation and Mann–Kendall significance analysis. Finally, an XGBoost machine learning model and the SHAP interpretable model are used to analyze the relationships between the driving factors and WUE. The findings of this study are significant for ecological civilization construction and sustainable development in the region.

2. Dataset and Study Area

2.1. Dataset

This study uses eight types of data (

Table 1. Data information.

Name	Unit	Source	Resolution	Timeframes
GPP	gC m−2	PML_V2	Day/500 m	2003–2020
ET	mm/d	PML_V2	Day/500 m
NDVI	-	USGS	16 Day/500 m
VPD	kPa	TerraClimate	Month/4 km
PDSI	-	TerraClimate	Month/4 km
SRAD	W/m2	TerraClimate	Month/4 km
TMAX	°C	TerraClimate	Month/4 km
PRE	mm	TerraClimate	Month/4 km

), all from the growing season (March to October), including ET and GPP data required for WUE estimation, as well as related driving factor data. These driving factors include the normalized difference vegetation index (NDVI), vapor pressure deficit (VPD), shortwave radiation (SRAD), maximum temperature (TMAX), the Palmer drought severity index (PDSI), and precipitation (PRE) data. ET and GPP data are sourced from the Penman–Monteith–Leuning Evapotranspiration V2 (PML_V2) (poles.tpdc.ac.cn, accessed on 8 July 2024), with a spatial resolution of 500 m and a temporal resolution of 8 days. The NDVI is sourced from the United States Geological Survey (USGS, lpdaac.usgs.gov, accessed on 8 July 2024), with a spatial resolution of 500 m and a temporal resolution of 16 days. Other driving factor data are from TerraClimate (www.climatologylab.org, accessed on 8 July 2024), with a spatial resolution of 4 km and a temporal resolution of 1 month. The annual GPP and ET data are obtained by summing the data, while the annual driving factor data are obtained by averaging data. Due to the different data sources, the spatial resolutions vary. In this study, a 500 m resolution is used for analyzing WUE. However, for the driver analysis, the WUE data are resampled to 4 km resolution to ensure consistency with the resolution of the driving factors, thereby maintaining uniformity in resolution. The study period spans from 2003 to 2020.

2.2. Study Area

The Yunnan–Guizhou Plateau in China is located in the subtropical monsoon climate zone of southern China (95°–110°E, 22°–30°N). It is a world-renowned karst landscape region and one of China’s four major plateaus (Figure 1a). The elevation ranges from 2 to 4 km (Figure 1b) and features a complex array of land use types (Figure 1c). The region has a rich and diverse vegetation environment, fostering abundant biodiversity and unique cultural diversity (Figure 1d). It belongs to a subtropical humid zone, with a climate strongly influenced by monsoons, resulting in significant climatic variation within the region. The Yunnan–Guizhou Plateau is built upon a widespread base of thick limestone, which, through surface and groundwater dissolution, has formed various typical karst landforms, including waterfalls, sinkholes, depressions, underground rivers, caves, gorges, natural bridges, and basins, making it one of the most developed karst regions in the world. Due to the severe issue of rocky desertification, ecological problems have long been a focal point for scholars and the government, becoming a key issue in promoting ecological civilization construction in the region.

3. Methods

The research content can be divided into three parts: data acquisition and processing, spatiotemporal evolution analysis, and driving factor analysis (Figure 2). All data processing for this study, including tasks such as stitching, clipping, resampling, and downloading, was performed on the Google Earth Engine (GEE) platform. Due to the varying resolutions of the data, a 500 m WUE dataset was used for the spatiotemporal evolution analysis, and for the driving factor analysis, the WUE resolution was resampled to match the resolution of the driving data, which is 4 km. The Theil–Sen slope estimator and Mann–Kendall significance analysis were then used to explore the temporal trends and spatial significance of WUE in the study area. Finally, XGBoost was employed to establish the relationship between the driving factors, and the SHAP model was used to interpret the importance of these driving factors and their specific relationship with WUE.

3.1. WUE Estimation

The estimation of WUE uses the commonly applied ratio method, and its calculation formula is as follows:

$$\mathrm{W}\mathrm{U}\mathrm{E}=\mathrm{G}\mathrm{P}\mathrm{P}/\mathrm{E}\mathrm{T}$$

(1)

where WUE represents water use efficiency, GPP is the gross primary productivity, and ET is evapotranspiration.

3.2. Estimation of Theil–Sen Slope

This method is a non-parametric method for slope estimation, and its advantage is that it is not affected by a small number of outliers. It primarily focuses on the changes in data over time [27]. It is used to estimate the variation trend of WUE in the study area.

$$\mathrm{S}\mathrm{L}=h\left({\displaystyle \frac{{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{a}\mathrm{f}}-{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{b}\mathrm{e}}}{\mathrm{a}\mathrm{f}-\mathrm{b}\mathrm{f}}}\right),\forall \mathrm{a}\mathrm{f}>\mathrm{b}\mathrm{f}$$

(2)

where SL is the slope, 0 < bf < af < L, and L represents the length of the time series. af and bf represent the values for the af-th year and the bf-th year, respectively. h is the function for calculating the median.

3.3. Mann–Kendall Significance Analysis

The Mann–Kendall method is a technique used to test the significance of temporal changes. It does not depend on the overall distribution of the data and can handle time series data with missing values [28].

$$\mathrm{S}\mathrm{V}={\sum }_{\mathrm{a}\mathrm{f}=1}^{l-1}{\sum }_{b\mathrm{f}=\mathrm{a}\mathrm{f}+1}^l\mathrm{s}\mathrm{g}\mathrm{n}({WUE}_{\mathrm{a}\mathrm{f}}-{WUE}_{\mathrm{b}\mathrm{f}})$$

(3)

$$\mathrm{s}\mathrm{g}\mathrm{n}\left({\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{a}\mathrm{f}}-{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{b}\mathrm{f}}\right)=\left\{\begin{array}{c}1,{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{a}\mathrm{f}}-{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{b}\mathrm{f}}>0\\ 0,{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{a}\mathrm{f}}-{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{b}\mathrm{f}}=0\\ -1,{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{a}\mathrm{f}}-{\mathrm{W}\mathrm{U}\mathrm{E}}_{\mathrm{b}\mathrm{f}}<0\end{array}\right.$$

(4)

$$\mathrm{S}\mathrm{I}=\left\{\begin{array}{c}{\displaystyle \frac{\mathrm{S}\mathrm{V}-1}{\sqrt{\mathrm{\phi }\left(\mathrm{S}\mathrm{V}\right)}}},SV>0\\ 0,\mathrm{SV}=0\\ {\displaystyle \frac{SV+1}{\sqrt{\mathrm{\phi }\left(\mathrm{S}\mathrm{V}\right)}}},\mathrm{SV}<0\end{array}\right.$$

(5)

$$\mathrm{\phi }\left(\mathrm{S}\mathrm{V}\right)=(l\left(l-1\right)\left(2l+5\right)-{\sum }_{r=1}^R{\mathrm{W}\mathrm{U}\mathrm{E}}_r\left({\mathrm{W}\mathrm{U}\mathrm{E}}_r-1\right)(2{\mathrm{W}\mathrm{U}\mathrm{E}}_r+5))∕18$$

(6)

where l represents the length of the time series, af and bf are the indices of the variables, R is the number of repeated data groups in the sequence, and r is the number of repeated data. S is the test statistic, SI represents significance, and $\mathrm{\phi }\left(\mathrm{S}\mathrm{V}\right)$ is the variance in SV. If the absolute value of SI is greater than 1.96, it indicates a significant change; otherwise, it indicates no significant change.

3.4. XGBoost Model

The XGBoost model is a representative algorithm of the boosting method in ensemble algorithms. By incorporating a regularization term into the objective function, it significantly enhances the generalization ability of the model and helps prevent overfitting. Compared to traditional algorithms, it delivers superior performance in regression tasks. To eliminate errors caused by the disparity in the magnitudes or values of the driving factors, all driving factors in this study are normalized.

$${\widehat{y}}_i={\sum }_l^L{f}_l\left(X_i\right)$$

(7)

$$f_l\left(X_i\right)={\theta }_l·\mathrm{\beth }(X_i\in R_l)$$

(8)

where ${\widehat{y}}_i$ is the regression value of the model, L is the total number of trees, $f_l\left(X_i\right)$ is the regression value of the l-th tree for the predictor X_i, θ is the leaf node value, R is the leaf node region, and $\beth$ is the indicator function.

3.5. SHAP Model

To clarify the synergistic relationships among the factors influencing WUE, this study uses the SHAP model to explain the potential nonlinear relationships between WUE and six driving factors.

$${\varphi }_j={\sum }_{S\subseteq n\left\{j\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}}[f\left(S\cup \left\{j\right\}\right)-f(S)]$$

(9)

where ϕ_j is the SHAP value of feature j, f(S) is the model prediction on the feature subset S, and N is the set of all features.

4. Results

4.1. Spatiotemporal Evolution Characteristics

GPP and ET are important indicators for estimating WUE and are commonly used vegetation indices to reflect vegetation growth status. Therefore, this study first presents the annual average spatial distribution of (Figure 3a) and temporal variation (Figure 3b) in GPP, as well as the annual average spatial distribution of (Figure 3c) and temporal variation (Figure 3d) in ET for the Yunnan–Guizhou Plateau. The spatial distribution of GPP shows significant variation, with high values mainly concentrated in the southwestern and southeastern regions, and low values in the northwestern and central regions. The range of values is primarily between 600 and 1500 gC m². The spatial distribution of ET is similar to that of GPP, with high values in the southwestern region and lower values in the northwestern and northeastern regions, with values ranging from 400 to 1000 mm. The northwest region, characterized by higher elevations, has limited vegetation growth, while the southwest region is more vegetated. Therefore, this spatial distribution is closely related to the local climate conditions and vegetation types. Additionally, from the temporal variation in GPP and ET, both show an increasing trend, with the growth rate of GPP at 5.40 and that of ET at 2.39. It is evident that although both are increasing, the growth rate of GPP is significantly faster than that of ET.

Figure 4 shows the temporal variation in (a), annual average spatial distribution of (b), and trend of change (c) in WUE in the study area from 2003 to 2020. From the temporal variation in WUE, there is an overall increasing trend, with a growth rate of 0.002. The WUE value is lowest in 2012, at approximately 1.91 gC m², while it reaches the highest value of 2.15 gC m² in 2013. This indicates an improvement in the water absorption and utilization capacity of vegetation in the study area. The spatial distribution of WUE follows a “high–low–high” pattern from southwest to northeast, with lower WUE values in the northwest and central regions. The WUE values mainly range from 1 to 2.5 gC m². Additionally, the Theil–Sen slope method was used to estimate the areas of WUE increase and decrease, and corresponding regional distribution maps were created. The results show a decreasing trend in WUE in the western and southeastern regions, while other areas exhibit an increasing trend. This is because the karst landforms are more prominent in the central region, and the higher phosphorus concentration and leaf thickness in the karst ecosystem lead to a higher photosynthetic rate, resulting in a higher WUE in these areas compared to others. Notably, the region with higher WUE in the southwest shows a decreasing trend, while the area with lower WUE in the central region exhibits an increasing trend. This phenomenon suggests that the WUE variation trends in different regions are closely related to local climate, vegetation, and ecological changes, potentially influenced by multiple factors.

Next, based on the Theil–Sen slope results and combined with the Mann–Kendall significance analysis method, the spatial distribution map of significant changes in WUE in the study area was generated (Figure 5a), and the area proportions of each change type were calculated (Figure 5b). From the spatial distribution map of significant changes, it can be seen that WUE changes in the study area exhibit four types: significant increase, weak increase, weak decrease, and significant decrease. The regions with a significant increase are mainly concentrated in the central area, covering 6.68% of the total area; other WUE increase areas are predominantly characterized by weak increases, which account for 50.80% of the area. The regions with a significant decrease are primarily located in the southwest, covering 4.60% of the area, while the remaining WUE decrease areas show weak decreases, accounting for 37.92% of the area.

4.2. Driving Factor Analysis

This study employed the SHAP model to calculate the SHAP values of each driving factor and created a summary plot of driver importance (Figure 6). In terms of importance ranking, the driving factors are ordered as follows: NDVI, TMAX, SRAD, PDSI, VPD, and PRE. Among them, the SHAP values of the NDVI, TMAX, and VPD increase from negative to positive as their feature values shift from low (blue) to high (red), indicating that these three factors generally promote an increase in WUE. On the other hand, the SHAP values of SRAD and the PDSI shift from negative to positive as their feature values decrease, suggesting that these two factors generally suppress WUE growth. For PRE, the changes in feature values are more complex, indicating that there may be a non-monotonic relationship between PRE and WUE or that its influence may not be statistically significant. This analysis reveals the different mechanisms through which various drivers influence WUE changes. Specifically, changes in the NDVI, TMAX, and VPD tend to promote an enhancement in WUE, while changes in SRAD and the PDSI may somewhat limit its growth.

Based on the SHAP values from the model, this study further examined the driving relationships between each factor and WUE (Figure 7). The results show that the NDVI and SRAD have monotonic effects on WUE, exhibiting linear and quadratic polynomial relationships, respectively. Specifically, the NDVI has a linear, monotonically increasing effect on WUE, with a slope of 0.87, meaning that WUE consistently increases as the NDVI rises. In contrast, SRAD exhibits a monotonically decreasing quadratic relationship with WUE, where the rate of decline in WUE decreases as SRAD increases. This suggests that the negative impact of high solar radiation on WUE is mitigated at higher SRAD values, possibly due to the influence of other factors. The effects of TMAX, the PDSI, VPD, and PRE on WUE are more complex, showing an initial increase followed by a decrease. For instance, the relationship between TMAX and WUE follows a quadratic pattern, indicating that while higher temperatures may initially enhance WUE, their effect becomes inhibitory as temperatures continue to rise. Similarly, precipitation (PRE) also follows a quadratic curve, suggesting that moderate levels of precipitation boost WUE, while excessive amounts may reduce it. The relationships for the PDSI and VPD are more intricate, following cubic polynomial trends, which indicates that the impacts of drought conditions and vapor pressure deficit on WUE vary non-linearly under different climatic circumstances.

5. Discussion

5.1. Spatiotemporal Evolution Characteristics

This study estimates the WUE of the Yunnan–Guizhou Plateau in China using the commonly applied GPP-to-ET ratio method. The results show a significant increasing trend in WUE from 2003 to 2020 in this region. Similar conclusions were drawn by Song et al. [29] in their analysis of WUE changes from 2009 to 2013. Furthermore, this study also reveals the spatial distribution characteristics of WUE, which follow a “high–low–high” pattern from the southwest to the northeast. Specifically, WUE is relatively lower in the northwest and central regions, whereas the southeast and eastern regions exhibit significantly higher WUE. This spatial variation is closely related to the distribution of vegetation types in the study area. Different vegetation types exhibit different WUE [30], which in turn influences the spatial variation of WUE in the region. In the southwestern part of the region, forest predominates, while other areas are primarily covered by grasslands. Forest, particularly the leaf growth processes in forested areas, is more conducive to water absorption and utilization, leading to higher WUE values in this region. This finding is supported by Qu et al. [31], who found that WUE in forest-covered areas is generally higher than in grassland areas. This is likely due to the more efficient transpiration process of forests and their better utilization of water.

5.2. Analysis of Driving Factors

This study, based on the XGBoost model and SHAP value analysis, explored the impact of different driving factors on WUE. The results show that the NDVI, TMAX, and VPD have a significant positive effect on WUE, indicating that the improvement in WUE in the study area is closely linked to the contributions of these factors. This is because the increase in the NDVI, under the influence of temperature and VPD, enhances the ecosystem’s carbon sequestration capacity and improves WUE [32]. Notably, the effects of TMAX and VPD are more complex. TMAX exhibits a quadratic relationship with WUE, as moderate temperatures promote WUE, whereas excessively high temperatures can severely impact photosynthesis, stopping organic matter production and transpiration [33,34]. On the other hand, SRAD and the PDSI have a suppressive effect on WUE. High SRAD increases evaporation, which in turn reduces WUE [35]. The PDSI shows a suppressive effect on WUE because, in environments with low drought severity, vegetation can increase productivity, but in persistent drought conditions, vegetation struggles to absorb nutrients and grow [36]. In sum, the findings of this study highlight the intricate mechanisms through which different driving factors influence WUE, underscoring that the relationship between these factors and WUE is not merely linear. Additionally, given the complex impacts of temperature and VPD on WUE, regional climate change management should focus on how to reduce the negative effects of extreme temperatures and droughts on water resources through adaptive management. For instance, water resource reserves and improvements in irrigation technology can help cope with extreme drought conditions, thereby enhancing the adaptability of crops and natural ecosystems. Establishing a remote sensing-based monitoring system to track key factors in real time can provide early warnings for ecosystem management and help decision-makers take timely action in response to increasing climate change and environmental pressures.

5.3. Limitations

This study analyzed the impact of six driving factors on WUE using the XGBoost and SHAP models, revealing the nonlinear relationships between these factors. However, there are some limitations to the research. First, the study considered a relatively small set of driving factors. Future research could incorporate additional variables, such as soil moisture and land use changes, to enhance the model’s comprehensiveness. Second, the resolution of the driving factor data used in this study is relatively low. Future research should utilize higher-resolution meteorological and remote sensing data to more accurately capture changes at local scales. In addition, the analysis in this study is based on interannual data. In the future, subdividing the time span into quarterly or monthly intervals could indeed provide a deeper insight into the seasonal variations in WUE and the dynamic changes in its driving factors. These improvements would enhance the understanding of WUE variation mechanisms and provide more precise support for ecological and environmental management.

6. Conclusions

In order to explore the evolution characteristics of WUE and the driving factors influencing it in the Yunnan–Guizhou Plateau of China from 2003 to 2020, this study used Theil–Sen slope estimation and Mann–Kendall significance analysis to investigate the temporal trends and spatial significance of WUE changes in the study area. XGBoost was then used to establish the driving relationships, and the SHAP model was applied to explain the importance of the driving factors and their specific relationships with WUE.

(1) Overall, WUE showed an increasing trend, with a growth slope of 0.002, indicating that the vegetation’s water absorption and utilization capacity in the study area has improved and water resource management has made progress.

(2) The spatial distribution of WUE exhibits a “high–low–high” pattern from southwest to northeast, with lower WUE values in the northwest and central regions. Notably, the areas in the southwest with higher WUE showed a decreasing trend, while the areas in the central region with lower WUE showed an increasing trend.

(3) WUE showed a decreasing trend in the western and southeastern regions, while other regions displayed an increasing trend. Specifically, 6.68% of the area showed a significant increase in WUE, 50.80% showed a slight increase, 4.60% showed a significant decrease, and 37.92% showed a slight decrease. In the future, stronger water resource protection and management should be implemented in the western and southeastern regions.

(4) The importance of the driving factors is ranked as follows: NDVI, TMAX, SRAD, PDSI, VPD, and PRE. The NDVI exhibited a linear positive relationship with WUE. SRAD had a quadratic decreasing impact on WUE, and its inhibitory effect weakened at higher values. TMAX, PRE, the PDSI, and VPD all showed a non-monotonic relationship with WUE, exhibiting a pattern of increasing first and then decreasing. The patterns of these driving factors provide a theoretical basis for regional water resource management and can also serve as a reference for other regions with similar ecological conditions.