The GUB is situated in the upper reach of the Ganjiang River (Figure 1), in Jiangxi Province, southern China, encompassing 18 counties of Ganzhou city. It is situated between latitudes 24°29′N and 27°09′N, and longitudes 113°54′E and 116°38′E, encompassing a total area of 39,379.64 km², which constitutes 23.6% of the total area of Jiangxi Province. The region’s topography is predominantly characterised by mountainous and hilly terrain, which collectively constitute the geomorphological framework and account for 80.98% of the total area. The GUB has a subtropical monsoon climate because it is located on the southern border of the mid-subtropical zone. The predominant winter and summer monsoons, heavy spring and summer precipitation, clear seasonal variations, a moderate temperature, copious amounts of heat and precipitation, brief episodes of extreme cold or heat, and an extended frost-free season are its defining features. The average annual temperature in the area is 19.8°C, while the average annual precipitation is 1,318.9 mm (Xu, 2025).

The data comprised digital elevation model (DEM), land use data, Soil Organic Carbon (SOC) and grain output (Table 1). The grain output and GDP data were provided by the Ganzhou Municipal Bureau of Statistics (2021), while the product price data was sourced from the China Agricultural Products Price Survey Yearbook (2021). The research framework is shown in Figure 2. In order to standardize the spatial resolution of multi-source data, the present study employed resampling techniques to normalize the raw data. The study area was meticulously delineated into a regular grid of 5 km × 5 km cells. The mean value of all data points within each grid cell was calculated using spatial statistical methods, and these means were used as training samples for the machine learning model.

In this study, the equivalent factor method was employed to ascertain the value of individual ecosystem service equivalent factors and to estimate the coefficients for different land use types (Xie et al., 2003). According to data from the Ganzhou Statistical Yearbook, In 2020, the average grain production per unit area of the GUB was 9,792.71 kg/hm². Considering that the research duration extended over 3 decades, the fixed price of 2020 was utilized to correlate the ESV of each period. The principal crops cultivated in the GUB are rice, vegetables, and fungi, along with oil crops. The mean prices for rice, vegetables, fungi, and oil crops in 2020 were 3.53 yuan/kg, 2.29 yuan/kg, 4.7 yuan/kg, and 2.21 yuan/kg, respectively. Thus, in 2020, the economic value of the equivalent factor of the GUB was 4,109.01 yuan/hm², which allowed for the calculation of the ESV coefficients for the GUB (Table 2). The calculation of ESV is shown in Equations 1, 2 (Hu et al., 2023; Shi et al., 2012): E S V = ∑ A k × V C k (1) E S V k = ∑ A k × V C f k (2)

Where ESV represents the total value of ecosystem services, V C k is the ecosystem value coefficient, A k is the area of land use type k, E S V k is the value of the fth ecosystem service function, and V C f k is the value coefficient of the fth service function for land use type k.

Grids measuring 2 km by 2 km, 5 km by 5 km, 8 km by 8 km, and 10 km by 10 km were first built as potential assessment units for this study. The ArcGIS10.8 software tools, including Create Fishnet, Clip, and Dissolve, were employed to select a grid size of 5 km × 5 km, resulting in a total of 1,744 small grids. Based on the land use data available for each grid, the ESV (ESV) of each land use within each grid was calculated, and the values were then summed to obtain the total ESV for the grid. The following Equations 3, 4 provide the formula (Li et al., 2018; Yang et al., 2022): E S V n = ∑ S m n × L U m (3) E S V = ∑ n = 1 k E S V n (4)

ESV _n represents the ESV (ESV) of the nth land use type within each grid cell; S _mn denotes the area (in hectares) of the nth land use type in the mth grid cell; LU _m is the ESV coefficient for the mth land use type; and k is the total number of land use types.

The regional variability of ESV can be reflected by utilising the relative change rate, as expressed by the following Equation 5 (Liu G. et al., 2014): R = R L R C = L b − L a L a C b − C a C a (5)

R represents the relative change rate, where R _L and R _C are the regional and global change rates, respectively; L _a and L _b denote the initial and final values of regional ESV; and C _a and C _b represent the initial and final values of global ESV. The relative change rate exhibits the following characteristics:

(1) It has been demonstrated that when the relative change rate is greater than or equal to 0, regional ESV trends exhibit alignment with global trends. When the relative change rate is greater than or equal to 0 but less than 1, regional change is weaker. In contrast, a relative change rate that is greater than 1 indicates that regional change exceeds global change.

“Spatial autocorrelation” refers to both local and global types of spatial autocorrelation. The global Moran’s index (Moran’sI) is employed to delineate the prevailing trend of spatial correlation in ESV. The local spatial autocorrelation index, represented by Moran’sI _i , reflects the degree of spatial autocorrelation within grid cells (Pearson, 2002). The formula is presented in Equations 6-8 as follow (Jin et al., 2020; Yao et al., 2015): M o r a n ′ s I = n ∑ i = 1 n ∑ j = 1 n w i j x i − x ¯ x j − x ¯ ∑ i = 1 n ∑ j = 1 n w i j ∑ i = 1 n x i − x ¯ 2 (6) M o r a n ′ s I i = x i − x ¯ m 0 ∑ j = 1 n w i j x j − x ¯ (7) m ˙ 0 = ∑ i = 1 n X i − x ¯ 2 n (8)

Where n represents the number of spatial units; x i and x j denote the ESV values of the ith and jth spatial units, respectively; x is the mean ESV; w i j is the spatial weight matrix; and m ˙ 0 denotes the variance.

In this study, a 5 km × 5 km grid was employed to subdivide the study area into 1,744 grid cells, and the mean values of environmental variables within each grid were extracted as feature data. Subsequent to the collection of these data, five machine learning models are constructed, including eXtreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Random Forest (RF), HistGradientBoosting (HGB), and Linear Regression (LR), to predict ESV. Table 3 offers a synopsis of the fundamental assumptions and the applicable scopes of the five models utilized in this study for ESV prediction. The transition from linear benchmark (LR) to tree-based models facilitates a comprehensive comparison chain, thereby enabling the validation of nonlinear gains. The algorithms RF, GBM, XGBoost, and HGB do not prespecify functional forms, a property that enables them to capture nonlinearities, interactions, and threshold effects between ESV and multisource drivers. The XGBoost and HGB models have been demonstrated to be effective tools for the analysis of large, high-dimensional datasets, with the capacity to meet the demands of multi-scale ESV mapping (Chen and Guestrin, 2016; Ke et al., 2017; Zhang et al., 2020).

The dependent variable in this study was ESV, and the independent variables included six natural factors, elevation (Ele), slope, precipitation (Pre), temperature (Tem), soil organic carbon (SOC) and fractional vegetation cover (FVC), and two socioeconomic factors, human activity intensity (HAI) and gross domestic product (GDP). The distributions of the eight variation are shown in Figure 3 (The elevation is shown in Figure 1). To quantify the effect of each factor on ESV, the study incorporated the SHAP analysis method. The dataset was split into two parts randomly: a test set and a training set. The percentage of the former to the latter was 20% and 80%, respectively. The performance of the model was evaluated based on the coefficient of determination (R ²), the mean absolute error (MAE), and the root mean square error (RMSE). The optimal model is applied to the trained model for driving factor detection.

The extraction of elevation and slope data was conducted using DEM, while the calculation of the FVC and HAI indicators was informed by relevant literature.

The calculation of FVC are shown in Equation 9 (Li et al., 2021): F V C = N D V I − N D V I min N D V I max − N D V I min (9)

In the formula, FVC represents the vegetation cover (%), NDVI is the normalized difference vegetation index, with the multi-year monthly average used in this study. The NDVI value of pixels that are fully covered by vegetation is known as NDVI_max, whereas the NDVI value of bare soil or regions devoid of vegetation cover is known as NDVI_min.

The HAI was employed to quantify the intensity of human interference in a specific region’s landscape. The HAI value ranged from 0 to 1, with higher values indicating greater human activity impact on the landscape components. The formulas are shown in Equation 10 (Yan et al., 2014): H A I = ∑ i = 1 N A i P i T A (10)

Where, HAI is the comprehensive index of anthropogenic impacts, N is the number of landscape types, Ai is the area of the ith type of landscape, P _i is the intensity coefficient of anthropogenic impacts of the ith type of landscape, the P _i in this study were given values of 0.12, 0.61, 0.09, 0.12, 0.94, and 0.08 for forestland, farmland, grassland, water body, construction land, and unused land, respectively, based on results of other studies (Yan et al., 2014). TA is the total area of the landscape.

In this study, Pearson’s correlation coefficient was employed to evaluate the relationship between each factor and SHAP values. In order to more accurately depict the nonlinear relationship between the eigenfactors and SHAP values, the LOWESS (locally weighted scatter smoothing) method was introduced to smooth the data. SHAP (SHapley Additive exPlanations) is a game theory-based method for interpreting machine learning models. The methodology employed entails the quantification of the marginal contribution of each input feature variable to the prediction results of individual samples through the utilization of Shapley values (Dandolo et al., 2023). This approach not only elucidates the trend of model predictors but also underscores the significance of features. The following mathematical expressions in Equation 11 are pertinent to this discussion (Aas et al., 2021): y i = y b a s e + f x i , 1 + f x i , 2 + ⋯ + f x i , j + ⋯ f x i , k (11)

In this equation, y_i denotes the predicted ESV value of the ith sample, y_base signifies the mean ESV prediction of all samples, x_i represents the observed ESV value of the ith sample, f (x_i,j) is the SHAP value of the jth feature of the ith sample, and k denotes the number of input features.

In this study, the Bootstrap method is employed to analyze the uncertainty in the prediction results of SHAP values. Specifically, the generation of multiple Bootstrap samples is achieved through the implementation of 100 resampling with put-back from the original sample. For each Bootstrap sample, the corresponding statistic is computed, thereby estimating the probability distribution of the original statistic. Consequently, we are equipped to derive confidence intervals for the raw statistic. In this context, 95% confidence intervals are delineated by referencing the 2.5% and 97.5% quantiles of the Bootstrap sample statistic. The width of the confidence interval is a critical metric for evaluating the precision of the estimate. Narrower confidence intervals are indicative of precise estimates, while wider intervals suggest greater uncertainty (Sothe et al., 2022). The Bootstrap method has been demonstrated to effectively quantify the uncertainty of the prediction results of SHAP values, thereby providing a more reliable basis for model interpretive analysis.

As demonstrated in the following section, partial dependence plots facilitate a more thorough elucidation of the specific association patterns between each feature and the target variable. The intersection of the dashed line (SHAP = 0) with the LOWESS trend line delineates the SHAP values into two regimes: values above the dashed line (SHAP >0) indicate a positive influence of the driver on ESV, whereas values below the dashed line (SHAP <0) denote an inhibitory effect (Lu et al., 2025).

Figure 4 illustrates the distribution of landscape types in GUB in 1990 (a), 2000 (b), 2010 (c), and 2020 (d). Table 4 presents a summary of the observed changes in these landscape types. The findings indicated that forestland has remained the predominant landscape type in GUB, encompassing over 74% of the total area. From 1990 to 2020, the area of forestland exhibited a reduction of 388.19 km². The second-largest landscape type, farmland, exhibited relatively stable conditions. Even while construction land made up a very small percentage of the total area (less than 2%), this category had grown significantly, with a net gain of 363.97 km², or 105.21%. The area of water bodies increased by 12.86 km², while the area of grassland increased by 15.01 km². Unused land exhibited a general decreasing trend, with a small proportion throughout the study period, remaining consistently below 0.1%.

The natural breaks method was utilised for the classification of the ESV index (ESVI) across the various grids into five distinct categories. The categories are ordered from low to high as follows: The categories are defined as follows: low (ESVI ≤24000 yuan/hm²), lower (24000 yuan/hm² < ESVI ≤48000 yuan/hm²), central (48000 yuan/hm² < ESVI ≤64000 yuan/hm²), higher (64000 yuan/hm² < ESVI ≤88000 yuan/hm²), and high (88000 yuan/hm² < ESVI). The resulting classification produced spatial variation maps of ESV of the GUB (see Figure 6). The distribution exhibited a distinctive pattern, with elevated values observed at the county boundaries and diminished values at the county centres. The boundaries of Shangyou County and Chongyi County were the main locations with the greatest ESV levels. In contrast, areas with low and relatively low ESV were predominantly situated in Ruijin City, Shicheng County, Nankang District and Xinfeng County, Dayu County.

Comparing the ESV levels at two different times shows that a “grade stability zone” is created if the two ESV levels are the same. The latter is referred to as a “grade up zone” if its ESV level shows improvement. Lastly, a “grade drop zone” is created if the level of the subsequent ESV declines. The results indicate that from 1990 to 2000, 95.82% of the area exhibited stable ecosystem service value (ESV) levels, while 3.42% of the area demonstrated an increase in ESV levels. Conversely, only 0.76% of the area exhibited a decrease in ESV levels, representing a very small proportion and distributed sporadically (Figure 7). From 2000 to 2010, 89.88% of the area exhibited stable ecosystem service levels, while 0.44% experienced an increase, which was minimal and distributed sporadically. Nevertheless, a decrease in levels was observed in 9.68% of the area. From 2010 to 2020, 95.66% of the area exhibited stable ESV levels, with only 0.17% demonstrating an increase, representing a minimal proportion and sporadic distribution. Conversely, a decrease in levels was observed in 4.16% of the area. From 1990 to 2020, 88.13% of the area exhibited stable ESV levels, with 0.69% demonstrating an increase. Conversely, 11.19% of the area exhibited a decline in ecosystem service levels, with the greatest concentration of affected areas observed in Dayu County, Shangyou County, Zhanggong District, and Nankang District. From a broader perspective, the proportion of areas with decreasing ESV levels of the GUB is significantly higher than those with increasing levels, indicating a risk of decline for regional ESV.

This study selects 18 counties or districts of the GUB as the subjects of analysis, and the differences in ESV among districts/counties are presented in Table 6. During study period, the ESV of Shangyou and Ganxian exhibited an increase of 0.21% and 0.67%, respectively, while the ESV in the remaining 16 counties or districts demonstrated a decline. Zhanggong exhibited the most significant decline, at 8.44%, with Nankang, Dayu, Longnan, and Dingnan also demonstrating declines exceeding 1%.

To analyse regional variability using relative change rates (R), it was observed that from 1990 to 2000, the ESV of the GUB increased by 0.12%. The R values for three counties (Ganxian, Shangyou, Xunwu) exceeded 1, indicating that their ESV increased at a rate greater than that observed for the GUB as a whole. From 2000 to 2010, the ESV of the GUB exhibited a decrease of 0.49%. R values larger than 0 were shown in all 18 counties, suggesting that the ESV changes there were consistent with the general trend of the GUB. From 2010 to 2020, the ESV decreased by 0.47%. Once more, all counties exhibited R values greater than 0, thereby confirming the alignment of ESV changes with the overall trend.

From 1990 to 2020, the ESV of the GUB exhibited a net decline of 0.84%. The R values for Shangyou and Ganxian were less than 0, indicating an increase in these two counties, while the remaining 16 counties experienced a reduction in ESV, consistent with the findings of previous change rate analyses. It is noteworthy that six counties/districts (Zhanggong, Nankang, Dayu, Longnan, Dingnan, Yudu) exhibited R values greater than 1, indicating that these areas experienced greater ESV changes than the overall region, with similar trends.

To quantify the effect of each factor on ESV, the study integrated machine learning model with the SHAP analysis method. The performance of the model was evaluated based on the R ², MAE and RMSE. Due to the substantial computational demands of this approach, only the data of 2020 were used to identify the driving factors of ESV.

The model evaluation metrics are delineated in Table 7, and the optimal model is identified through comprehensive analysis in terms of performance, error, and risk of overfitting. R ² (Test) is a measure of the explanatory power of the model on the test set, and the closer it is to 1, the better the model performance is. The findings indicate that HGB (0.62) surpasses XGBoost (0.59), RF (0.55), GBM (0.53) and LR (0.41). Consequently, HGB emerges as the most effective model. MAE (Test) denotes the Mean Absolute Error, with smaller values being preferable. The resultant data indicates that HGB (1905.28) is outperformed by XGBoost (1965.46), RF (1978.37), GBM (2079.60) and LR (2540.72), with HGB demonstrating the lowest error rate. RMSE (Test) is the Root Mean Squared Error, and a smaller value is preferable. The findings indicate that HGB (3229.46) is outperformed by XGBoost (3351.19), followed by RF (3499.00), GBM (3555.46), and LR (3996.22), with the lowest HGB error. It is evident that HGB demonstrates superior overall performance on the test set, as evidenced by its elevated R ², minimal MAE, and RMSE. The negligible discrepancy between Train (0.84) and Test (0.62) further substantiates its reduced risk of overfitting and enhanced generalisation capability. Consequently, the optimal model is identified as HGB, while the alternative model is XGBoost.

The HGB plus SHAP analysis revealed the extent and direction of the effects of each environmental factor on ESV. The eight impact factors were then classified into three distinct grades, based on the results of the feature importance calculation (Figure 10). These three grades were designated as key factors, relatively important factors, and low-impact factors, respectively. Key factors are defined as those with feature importance ≥2000, Human Activity Intensity (HAI) emerged as the key factor, with a contribution value of 2285.21 (30.57%), exhibiting a substantial negative effect. That is, elevated HAI values were associated with diminished ESV predictions. The relatively important factors are those with feature importance of 500 ≤ importance value <2000, and include FVC (1681.43, 22.50%, negative effect), elevation (1038.79, 13.90%, negative effect) and SOC (851.48, 11.39%, positive effect). The low-impact factor, which is equivalent to the importance value of <500, encompasses the GDP (461.25, 6.17%, nonlinear relationship), slope (455.53, 6.09%, nonlinear relationship), temperature (362.35, 4.85%, nonlinear relationship) and precipitation (338.55, 4.53%, nonlinear relationship). Essentially, it was determined that the main parameters influencing changes in ESV were HAI, FVC, Ele, and SOC. In contrast, the contributions of economic, topographic and climatic factors were found to be comparatively negligible, though their influence should not be overlooked.

The present study investigated the effects of HAI, FVC, Ele, and SOC on ESV in the study area. The findings indicated that these factors exerted substantial influence on ESV, with divergent mechanisms of action being observed. The contribution of the key characterization factors to the model predictions was quantified by SHAP feature dependency plot analysis, and the nonlinear relationships and thresholds between the features and the predictions were revealed. Figure 11 illustrates these relationships.

A highly significant negative correlation was identified between HAI and SHAP values (r = −0.78, p < 0.001). As the HAI level increased, its effect on ESV shifted from promotion to inhibition. When HAI was less than 0.23, it exhibited a positive effect on ESV; conversely, when HAI was greater than 0.23, it demonstrated a negative effect.

A highly significant negative correlation was also identified between FVC and SHAP values (r = −0.92, p < 0.001). When FVC was less than 0.76, it exhibited a promoting effect on ESV; conversely, when FVC was greater than 0.76, the effect changed to inhibitory.

A highly significant negative correlation was also identified between altitude and SHAP values (r = −0.89, p < 0.001). At altitudes less than 400 meters, ESV exhibited a promoting effect, while at altitudes greater than 400 meters, the effect transitioned to inhibition.

A highly significant positive correlation was identified between the SHAP values and the SOC (r = 0.77, p < 0.001). When the SOC was less than 14 g/kg, it exhibited an inhibitory effect on ESV. Conversely, when the SOC was greater than 14 g/kg, the effect changed to a promoting one. The SHAP values exhibited a marked increase with rising SOC levels. However, this increase began to plateau after reaching a maximum value when the SOC level exceeded 20 g/kg.

In summary, the findings of the study demonstrated a non-linear, highly significant negative correlation between HAI, FVC, and elevation with ESV in the study area. Conversely, SOC exhibited a non-linear, highly significant positive correlation with ESV.

A thorough analysis of the uncertainty inherent in the predicted results of ESV was conducted in this study. To this end, we employed a 95% confidence interval (CI) to quantify the uncertainty. To illustrate this uncertainty, gray shading was employed to represent the 95% CI. The width of the shading is proportional to the size of the 95% CI; that is, the wider the shading, the greater the uncertainty in the prediction results.

The 95% CI of the ESV prediction result was found to be significantly larger when the HAI was greater than 0.6. This finding suggests that the ESV prediction results exhibit enhanced accuracy within the range of HAI less than 0.6. The 95% CI of the ESV prediction results was found to be significantly greater when the FVC was less than 0.6. This finding suggests that ESV predictions are more precise in cases where FVC is greater than 0.6. The 95% CI of the ESV prediction results exhibited a substantial increase when Ele was greater than 1000 meters. This finding suggests that ESV predictions exhibited higher accuracy within the range of elevation less than 1000 meters. The 95% CI of the ESV prediction results was found to be significantly greater when the SOC was less than 10 g/kg or greater than 25 g/kg. This finding indicates that ESV was more precisely predicted in the range of 10 g/kg ≤ SOC ≤25 g/kg.

In summary, the analysis of the 95% CI enabled the identification of key thresholds affecting the accuracy of the ESV prediction and the optimization of the predictive performance of the model accordingly. This finding is significant for improving the accuracy and reliability of ecosystem service value assessment.

In order to quantify the effect of each factor on ESV accurately, an integrated explainable machine learning approach was conducted in this study. The results indicate that HGB demonstrates superior overall performance on the test set, as evidenced by its elevated R ², minimal MAE, and RMSE. The HGB model demonstrates distinct advantages over linear regression approaches by effectively capturing nonlinearity without requiring linearity assumptions. Compared to RF model, it is more computationally efficient and more adapted for analyzing huge data sets. Compared to XGBoost or GBM model, the HGB model has the benefits of seamless integration with Scikit-learn, and compatibility with sparse data (Du et al., 2025). HGB is an advanced machine learning method designed to optimize the equilibrium of efficiency, accuracy, and user-friendliness in identifying feature importance. It is an effective approach for driving factor selection (Bridge et al., 2025).

The study indicated that HAI is the key driving factor of ESV of the GUB, with a feature important of 30.57%, exhibiting a substantial negative effect of pearson analysis. SHAP-based non-linear analysis revealed that when HAI was less than 0.23, it exhibited a positive effect on ESV; conversely, when HAI was greater than 0.23, it demonstrated a negative effect. In the formula of HAI, construction land had the greatest human impact intensity coefficients among the six land use classifications. A net loss of 388.19 km²of forestland over the research period and a dramatic 105% increase in building land as a result of fast urbanization account for this decline in ESV. The rapid urbanization of GUB over the past 3 decades has resulted in a notable decline in agricultural and forestland and a notable rise in building area. The conversion of industrial structures, particularly the annual expansion of aquaculture regions, has resulted in the loss of some cropland (Chen et al., 2015; Wu et al., 2022). It is therefore evident that changes in land use of the GUB directly impact the ESV. The impact of anthropogenic activities on landscape patterns has resulted in a reduction in the ecosystem service capacity of the GUB (Qiao et al., 2025; Liu et al., 2022). This finding aligns with the results of previous studies in this field (Yao et al., 2025). The HAI is the predominant factor in ESV spatial differentiation, underscoring the repercussions of recurrent human activities on the ecological environment. The formulation of strict environmental protection policies and measures is imperative to enhance the effective management of human activities and to restrict unreasonable land use and development, thereby reducing the negative impact of human activities on the ecological environment (Hu et al., 2023).

A highly significant negative correlation was also identified between FVC and SHAP values in this study. When FVC was less than 0.76, it exhibited a promoting effect on ESV; conversely, when FVC was greater than 0.76, the effect changed to inhibitory. This finding is at odds with the conclusions of numerous preceding studies (Princelyn et al., 2025). This “abnormal” phenomenon is primarily attributable to the unique natural and socio-economic context of the GUB. A potential explanation for this phenomenon is that Ganzhou City had implemented large-scale artificial afforestation projects (primarily pure stands of cypress and pine) in rare earth tailings areas, slope farmland, and urban peripheries over the past 2 decades. These projects had led to a rapid increase in regional FVC. These artificial forests consist of a single tree species, exhibit simple community structures, and possess low understory biodiversity (Zhang et al., 2025). Consequently, the core ecological functions of carbon sequestration, water conservation, and soil retention are significantly lower than those observed in natural evergreen broadleaf forests. This phenomenon has resulted in a decline in ESV per unit area. Another potential explanation pertains to remote sensing NDVI saturation. High FVC areas (>0.7) are predominantly artificial pure forests, where NDVI has reached or approached saturation. This renders it impossible to distinguish differences in forest stand quality, thereby creating the illusion of “high FVC-low ESV” (Li Y. et al., 2025).

This study employs both grid-based and administrative scales to evaluate the total ESV, the unit area ESV, and their spatial distribution characteristics of the GUB. The grid method employed a 5 km × 5 km grid, representing a micro-scale approach, which facilitated the mapping of the spatial distribution of unit area ESV across the various levels within the study area. Furthermore, it permitted a comparison of the fluctuations in different levels at two distinct points in time. While micro-scale assessments can elucidate the distinctions between diverse ecological subsystems at larger scales, the outcomes obtained at smaller scales are frequently challenging to extrapolate for ecological regulation applications at broader scales (Zhang et al., 2022).

Conversely, the spatial distribution of ESV based on administrative divisions represents a macro-scale assessment, offering insights into the spatial distribution of ESV across the 18 counties or districts within the study area. By analysing relative changes, similarities and differences in the trends of each county can be identified in comparison to the overall region. Administrative scale assessments may be constrained by the influence of dominant land and ecological system components within the units, which may result in the omission of variations among subsystems. Nevertheless, the benefit of administrative scales lies in their practical applicability, enabling the collection of statistical data related to administrative units and the establishment of ecological management models, which are more straightforward to implement than grid-scale assessments (Huang et al., 2019; Yu et al., 2025). Therefore, it is advantageous to combine several scales to fully comprehend the ecological functional status of the assessment units. The study of scale characteristics in ecosystem service evaluations can be enhanced to improve the objectivity, credibility, and practicality of assessment results (Xu et al., 2019; Peng et al., 2017).

The ESV of the GUB was evaluated, revealing an overall decline in ESV from 1990 to 2020. To improve regional ecological service value, bolster ecological security, and advance the formulation of ecological compensation policies for water supplies, the following proposals and strategies are suggested:

(1) The major purpose is to improve approach aimed at conserving and rehabilitating forest ecosystems. Prioritizing the conservation of high ESV areas is essential. Development must be rigorously limited, and the safeguarding of natural forests along with the restoration of agricultural land to forest should be strengthened to improve climate and hydrological regulatory functions, which constitute 50 percent of the ESV (Sun et al., 2025). GUB boasts a substantial forest area, and through the implementation of afforestation and grassland restoration initiatives, it has attained a notable degree of greening in the southern Jiangxi region, approaching the upper limit of the achievable greening scale. Given the sensitivity of forest land to the ESV of the region, future research should prioritize the optimization of forest stand structure and the enhancement of forest quality within mountainous forest land. This approach is expected to contribute to the continuous enhancement of the climate regulation, biodiversity maintenance, soil conservation, and landscape aesthetic ecological service values of forest land. For forest reserves, conservation should be the primary focus, with the objective of maintaining the soil and water conservation and water source protection functions of forest land (Yang et al., 2024).