How to Achieve Integrated High Supply and a Balanced State of Ecosystem Service Bundles: A Case Study of Fujian Province, China

Abstract

Ecosystems are nonlinear systems that can shift between multiple stable states. Ecosystem service bundles (ESBs) integrate the supply and trade-offs of multiple services, yet the conditions for achieving high-supply and balanced states remain unclear from a nonlinear, threshold-based perspective. In this study, six representative ecosystem services in Fujian Province were quantified, and ESBs were identified using a Self-Organizing Map (SOM). By integrating the Multiclass Explainable Boosting Machine (MC-EBM) with the API interpretable algorithm, we propose a framework for exploring ESB driving mechanisms from a nonlinear, threshold-based perspective, addressing two key questions: (1) Which factors dominate ESB formation? (2) What thresholds of these factors promote high-supply, balanced ESBs? Results show that (i) the proportion of water bodies, distance to construction land, annual solar radiation, annual precipitation, population density, and GDP density are the primary driving factors; (ii) higher proportions of water bodies enhance and balance multiple services, whereas intensified human activities significantly reduce supply levels, and ESBs are highly sensitive to climatic variables; (iii) at the 1 km × 1 km grid scale, optimal threshold ranges of the dominant factors substantially increase the likelihood of forming high-supply, balanced ESBs. The MC-EBM effectively reveals ESB formation mechanisms, significantly outperforming multinomial logistic regression in predictive accuracy and demonstrating strong generalizability. The proposed approach provides methodological guidance for multi-service coordination across regions and scales. Corresponding land management strategies are also proposed, which deepen understanding of ESB formation and offer practical references for enhancing ecosystem service supply and reducing trade-offs.

1. Introduction

Ecosystem services (ESs) encompass the diverse benefits that humans derive from ecosystems [1]. Interactions among these services are complex [2], and enhancing a provisioning service often comes at the cost of multiple regulating services [3]. Ecosystems are inherently nonlinear, with multiple stable states separated by thresholds [4]. When driving factors exceed these thresholds, substantial declines in multiple ESs can occur. Globally, approximately 60% of ESs have degraded over the past century, threatening human well-being [5], with ineffective management being a major contributor [6,7]. Coordinating the management of multiple ESs to achieve both high supply and balanced states is therefore an urgent challenge in sustainability science [8].

However, most ESs research to date cannot truly address these critical challenges for the integrated management of multiple ESs because it has focused primarily on quantifying and mapping the delivery of only a few services [8] and the driving mechanisms behind the relationships between arbitrary pairs of them [1,9,10]. ESs are interrelated in complex, multivariate ways. Increasingly, ecosystem service bundles (ESBs) are used to identify spatial clusters of multiple ESs [11], enabling the analysis of multivariate interactions and overall supply [12] for integrated management of multiple ESs because ESBs address a group of ESs rather than a single ES and, to a certain extent, bridge the noncontinuity in the demarcation of management target areas [13]. Numerous studies have used ESBs to research the spatial distribution characteristics of ES interactions under various application scenarios, such as climate change [14,15], urban expansion [16], ecological restoration measures [17], and ecological agriculture management [18]. However, these studies primarily focus on delineating ESBs and offer limited insights into their formation mechanisms [19].

Existing research shows that ESB formation is jointly driven by natural and socioeconomic factors [19]. The former directly determines the biophysical processes that produce ecosystems, while the latter directly or indirectly affects ecosystems through human intervention [20]. These factors include climate, vegetation, and human activities [7]. However, these studies mainly employed linear models (such as multivariate logistic regression (MLR) [21], geographic detectors [19], and redundancy analysis [14]), unexplainable machine learning models (such as random forests) [7], or spatial statistics and mapping methods. These approaches can identify the important driving factors of ESBs, but they cannot effectively visualize the specific driving processes or explore the threshold effects [4]. The mechanisms behind ESB formation remain unclear, including in which threshold ranges the “natural-social economic” factors promote the high supply and balanced clustering of multiple services or exacerbate trade-offs among services, leading to low supply clustering; thus, they remain to be further studied.

Moreover, ecosystems are nonlinear systems. Applying linear models, which rely on prior assumptions, to establish relationships among internal variables may introduce biases [10]. In contrast, machine learning models trained on raw datasets excel at capturing these complex relationships and faithfully replicating real ecological processes [22]. Additionally, integrating visualization-express algorithms into machine learning enhances interpretability by visually depicting relationships among variables through dependency graphs and facilitating the analysis of threshold effects [23]. However, ESBs present a challenge due to the multi-categorical attribute, which means that they do not align with the typical requirements of interpretable machine learning models for dependent variables (continuous or binary) [24]. Consequently, identifying the dominant drivers of ESBs and determining optimal driving thresholds from a nonlinear perspective remains a complex task. This study addresses this challenge by integrating the API (Additive Post-Processing for Interpretability) algorithm into the MC-EBM (Multiclass Explainable Boosting Machine) model. The MC-EBM model extends machine learning to tackle multi-class classification problems [25]. With the integration of the API, MC-EBM gains interpretability features, enabling it to visualize the relationships between variables in a map. This can not only clarify the dominant drivers behind ESBs but also help in identifying the optimal thresholds for these drivers.

This paper aims to address the following three questions: (i) What are the primary drivers of ESBs? (ii) How do natural and socioeconomic factors influence ESBs? (iii) Do these drivers have optimal value ranges that enable multiple ESs in a region to exist in a state of high supply and balanced clustering? Using Fujian Province as a case study, we spatially quantifie six typical ESs (Grain Production (GP), Carbon Sequestration (CS), Biodiversity Conservation (BC), Outdoor Recreation (OR), Soil Conservation (SC), and Water Yield (WY)) for the year 2015. Self-organizing map (SOM) neural networks are utilized to classify these ESs into five ESBs. Subsequently, 17 potential influencing factors were comprehensively selected from human activities, land use types, climate, terrain, and habitat characteristics, establishing a “natural-social-economic” potential driving force index system. Finally, by integrating MC-EBM and API, a visualization-express and interpretable multiclass machine learning model is constructed to investigate the dominant factors and optimal threshold ranges affecting the likelihood of high supply and balanced clustering of ESs. This study aims to offer insights into spatial optimization and management of regional ESs.

2. Materials and Methods

2.1. Study Area

Fujian Province stands as one of China’s pioneering provinces selected for comprehensive pilot initiatives in ecological civilization construction. It is renowned for its significant ecological importance and abundant natural resources. However, the province faces escalating challenges from rapid urbanization, which pose significant threats to its ESs. Therefore, the scientifically coordinated management of multiple ESs is crucial. This study thus selects Fujian Province as its research focus.

Geographically, Fujian Province spans from 23°33′ to 28°20′ north latitude and 115°50′ to 120°40′ east longitude. It boasts the highest forest coverage rate in the country at 66.8%, with a strong capacity for carbon sequestration by plants. The province hosts 4703 species of higher plants, including 52 species under national key protection, indicating a high level of ecological diversity. The water systems are well-developed, with a river density of 0.1 km/km². Major rivers include the Minjiang, Jiulongjiang, and Jinjiang, with a total basin area of 60,992 km², possessing strong WY capabilities. Fujian Province is predominantly mountainous, covering over 80% of its area, thereby posing high risks of soil erosion. Its varied river valleys create abundant ecological tourism resources. However, agricultural land is limited, with per capita arable land less than half of the national average and total grain production at approximately one-third of the national average, necessitating focused attention on enhancing the ecosystem’s food production capabilities. Fujian’s climate exhibits marked regional differences [9]. The southeastern coastal regions experience a subtropical climate, while the northeastern, northern, and western parts have a mid-subtropical climate, leading to substantial variations in hydrothermal conditions across different climate zones. Figure 1 provides an overview map of Fujian Province.

2.2. Data Source and Processing

The data types used in this study include raster data, vector data, monitoring station data, and statistical data. All non-spatial data underwent spatialization processing and were standardized to the Krasovsky_1940_Albers coordinate system. All data were resampled uniformly at a 1 km × 1 km grid size, with a reference year of 2015. The data used in this study and their sources are as follows: Digital Elevation Model (DEM) data with a resolution of 30 × 30 m from the Geographic Spatial Data Cloud Platform “http://www.gscloud.cn/ (accessed on 3 September 2025)”; Land Use/Land Cover map with a resolution of 30 × 30 m, Normalized Difference Vegetation Index (NDVI) at 1 km × 1 km resolution [dataset] [26], Chinese Vegetation Type map at 1 km × 1 km resolution, Gross Domestic Product (GDP) [dataset] [27] and population [dataset] [28] spatial distribution data at 1 km × 1 km resolution from the Resource and Environment Science and Data Center “http://www.resdc.cn/ (accessed on 3 September 2025)”; Meteorological dataset including temperature, precipitation, and sunshine hours from the China Meteorological Data Service Center “http://data.cma.cn/ (accessed on 3 September 2025)”; and total grain production data for prefecture-level cities in Fujian Province from the Fujian Statistical Yearbook “https://tjj.fujian.gov.cn/ (accessed on 3 September 2025)”.

2.3. Methods

2.3.1. The Assessment of Ecosystem Services

Drawing from Fujian Province’s distinctive geographical features—namely, substantial food security challenges, extensive forest cover (notable carbon sequestration potential), abundant biomass and ecotourism resources, hilly terrain (with elevated soil erosion risks), and a dense river network (prominent WY capacity)—our study evaluated six key ESs: GP, CS, BC, OR, SC, and WY.

Referring to previous research methods, the calculation methods for each service are outlined below: GP denotes the food supply capability per unit area of cultivated land. The Vegetation Condition Index (VCI) functions as an indicator of crop growth status [29]. GP values for each grid are calculated by distributing the total yield of grain and economic crops from each prefecture-level city based on the proportion of VCI-covered cultivated land within the city’s total VCI-covered cultivated land [4]. CS measures the net carbon dioxide fixed during photosynthesis minus that respired by plants, evaluated using the Carnegie–Ames–Stanford Approach (CASA) model [30,31]. BC assesses biodiversity conservation and habitat quality using the Integrated Valuation of ESs and Trade-Offs (InVEST) model [32]. OR evaluates the proximity between humans and natural environments using the ESTIMAP-recreation method [33]. SC quantifies the soil conservation capacity of practices (vegetation cover or artificial surface treatments) on bare soil, assessed via the Revised Universal Soil Loss Equation (RUSLE) [34]. WY indicates the equilibrium between water supply (precipitation) and loss (evapotranspiration [35] and surface runoff [36]) within an ecosystem, evaluated through the water balance equation [37].

2.3.2. Partitioning of Ecosystem Service Bundles

Using the self-organizing map (SOM) neural network model [38], spatial clustering of six ESs in Fujian Province for the year 2015 was conducted to delineate ESBs. SOM is a non-supervised classification method that does not require a predetermined number of categories; instead, it automatically divides the dataset into different categories based on its characteristics. The principle of SOM is to maximize similarity within ESBs and maximize dissimilarity between ESBs [39], ensuring thorough categorization of the dataset. The Davies–Bouldin index [40] is used to measure the scientific validity of data partitioning into different ESBs: a smaller Davies–Bouldin index indicates higher intra-cluster consistency and greater inter-cluster heterogeneity, ensuring comprehensive partitioning based on dataset characteristics. Therefore, the number of ESBs corresponding to the minimum Davies–Bouldin index is selected as the optimal combination for clustering ESs [41].

2.3.3. Construction of the “Natural-Social-Economic” Latent Driver Factor System

Based on prior knowledge and limited data, this study constructs a “natural-social-economic” latent driver indicator system, which comprehensively considers five dimensions: human activities, land use types, climate, topography, and habitat characteristics. It explores their impact mechanisms on ESBs. The “natural-social-economic” indicator system is shown in

Table 1. Comprehensive index system of the “natural-socio economic” conditions.

Select Dimension	Potential Impact Factor	Unit
Human activities	GDP density	RMB ten thousand/km2
	population density	People/km2
	distance from construction land	km
	distance from highway	km
	distance from railway	km
Land management	the proportion of cultivated land	%
	the proportion of woodland	%
	the proportion of grassland	%
	the proportion of water bodies	%
	the proportion of construction land	%
Climate	annual mean temperature	°C
	annual total precipitation	mm
	Monthly average solar radiation	MJ/m2
Topography	altitude	m
	slope	°
Habitat characteristics	NDVI	/
	distance from water bodies	km

.

2.3.4. Analysis of the Impact of “Natural-Social-Economic” Factors on Ecosystem Service Bundles

(i)

Construction of Training Dataset

ESBs served as the dependent variables. Due to their multiclass nature, the data samples for each category were uneven, potentially affecting the accuracy of machine learning [42]. To address this issue, a balanced preprocessing of the dependent variable dataset was conducted before applying the MC-EBM model. This involved using a hybrid sampling method that combines the synthetic minority oversampling technique (SMOTE) [43] with Tomek Links [44]. The “natural-social-economic” potential driving force indicator system was used as potential explanatory variables to construct the “dependent variable-explanatory variable” dataset.

(ii)

Capturing Relationships between ESBs and “natural-social-economic” factors

The MC-EBM model was employed to explore the impact mechanisms of ‘natural-social-economic’ factors on ES clusters. The MC-EBM model [25] utilizes boosting techniques to learn from the dataset and uncover associations between variables. (Figure 2) illustrates the foundational framework of the Boosting algorithm [4]. Initially, the first base learner trains on the input dataset, encompassing both the dependent and explanatory variables. It computes the residual, representing the difference between the predicted and actual values of the dependent variable. Subsequently, each successive base learner utilizes the residual output from the preceding learner as its training target, continuing this process iteratively until meeting the stopping criteria (e.g., precision thresholds). Ultimately, the goal is to minimize the residual through successive iterations.

This study employed 10-fold cross-validation (k = 10), dividing the dataset into 10 mutually exclusive subsets, with 9 used for training and 1 for validation in each iteration, to ensure the robustness of the evaluation results. A random seed of 1 was set to reduce fluctuations caused by randomness. To achieve optimal performance of the MC-EBM model in uncovering the driving mechanisms of ecosystem service bundles (ESBs), Bayesian optimization was applied to search for the best hyperparameter combinations [45]. Bayesian optimization constructs a surrogate probabilistic model to approximate the objective function of MC-EBM (i.e., minimizing residuals), enabling the identification of hyperparameter configurations that most significantly improve model performance with the fewest trials. This approach is considerably more efficient than grid search or random search methods. The hyperparameters involved in MC-EBM and their effects are summarized in Appendix A (

Table A1. Hyperparameters of the MC-EBM model.

Hyperparameter	Description	Search Range	Optimal Value
n_estimators	Number of weak learners. A larger value generally improves performance but increases computation cost and may lead to overfitting.	50–1000	50
max_bins	Maximum number of bins for feature discretization, affecting model complexity and fitting ability.	32–256	256
learning_rate	Learning rate controlling the contribution of each tree to the final prediction. Too high may cause overfitting, too low may slow convergence.	0.01–0.3	0.1
min_samples_leaf	Minimum number of samples per leaf. Larger values help reduce overfitting risk.	1–20	2
max_leaves	Maximum number of leaves per weak learner, controlling tree complexity.	3–20	3
outer_bags	Number of outer bagging iterations, used to reduce variance and improve model stability.	4–20	8
validation_size	Proportion of validation set used to evaluate model performance during training and prevent overfitting.	0.1–0.3	0.15
early_stopping_rounds	Number of rounds for early stopping. Training stops when validation performance does not improve for consecutive rounds, preventing overfitting.	10–100	50

).

(iii)

Visualization and Threshold Analysis of Factor Impacts on ESBs

Incorporating the API algorithm [24] into the MC-EBM model enables visualization of the effects of driving factors on ecosystem service bundles (ESBs), facilitating threshold analysis. For each driving factor, MC-EBM + API learns a separate additive model for each ESB category and generates a corresponding shape function (contribution curve). The vertical axis of each curve represents the contribution of a specific value of the factor to the log-odds of predicting that particular bundle, relative to a baseline. A positive contribution indicates that the factor value increases the likelihood of the model predicting the bundle, whereas a negative contribution indicates a decrease in this likelihood.

As illustrated in Figure 3, assume that bundle 1 represents the ideal bundle, characterized by strong and relatively balanced ecosystem service supply, while bundle 4 represents the bundle to be avoided, characterized by weak service supply. When the contribution of a factor to the log-odds of the ideal bundle exceeds its contribution to the log-odds of the bundle to be avoided, the corresponding range of factor values defines the optimal threshold interval.

3. Results

3.1. Spatial Patterns of Ecosystem Services

The six ESs in Fujian Province all exhibit significant spatial heterogeneity. The distribution of GP values demonstrates a notable correlation with administrative regions (Figure 4). The highest GP values are recorded in Longyan City. Despite the extensive and clustered farmland in Nanping City and the southeastern coastal region of Fujian, the grain production capacity per unit area of farmland is relatively low, leading to lower GP values. High CS values are primarily observed in high-altitude areas (Figure 1 and Figure 4), with a distribution pattern showing elevated values in the western and northern regions compared to the eastern and southern regions (Figure 4). Within certain threshold ranges, carbon storage in forest vegetation is positively correlated with precipitation and negatively correlated with temperature [31,46]. High-altitude regions typically experience lower temperatures, and Fujian Province’s climate exhibits significant regional variation: western Fujian receives more precipitation compared to eastern Fujian, while northern Fujian has lower annual temperatures than southern Fujian. The highest BC and OR values are found near water bodies; however, areas with more grassland and fewer forests exhibit higher OR values, whereas regions with more forests and less grassland show higher BC values (Figure 4). This is due to the consistency of land use in BC and OR [1]. They share the same supporting land types, but different land types provide varying levels of support for these two services. High SC values are predominantly found in hilly areas, with SC values generally higher in northern Fujian and lower in southeastern Fujian. This is attributed to the steeper slopes in hilly areas, where northern Fujian, with its higher total annual precipitation and altitude, faces a greater risk of soil erosion compared to southern Fujian. WY values display a clear west-to-east gradient, with higher values in northwestern Fujian due to its abundant precipitation.

3.2. Spatial Patterns of Ecosystem Service Bundles

The Davies–Bouldin index reveals that the clustering effect is optimal when ESs are divided into five ESBs (Figure 5c). Accordingly, ESs in Fujian Province are categorized into five distinct ESBs based on spatial clustering characteristics: Bundle 1 (low grain supply), Bundle 2 (high soil conservation supply), Bundle 3 (balanced and high ESs supply), Bundle 4 (highest ESs supply), and Bundle 5 (lowest ESs supply). These ESBs vary significantly in area and spatial distribution (Figure 5a). Bundle 1, the most widely distributed, is primarily found in non-cultivated areas. Bundle 2 is predominantly located in northwestern Fujian, particularly in the Wuyi Mountains region of Nanping City. Bundle 3 spans the entire province with a scattered distribution that includes cultivated lands, which are neither overly dense nor extensively contiguous. Bundle 4 is mainly situated in central Fujian, predominantly in high-altitude regions. Bundle 5 is concentrated in the eastern and southern parts of Fujian Province, notably in the densely developed and cultivated areas of southeastern Fujian’s coastal region.

The supply of ESs across the five ESBs exhibits the following commonalities (Figure 5b): (i) The supply levels of BC and WY services consistently rank highest among the six services. This is attributed to the high forest coverage in Fujian Province, which provides more primary supportive land types for BC across all ESBs. Additionally, Fujian’s ample rainfall provides sufficient water resources for WY services, with the highest precipitation levels located in the northwestern part of the province, resulting in the highest WY service provision in Bundle 2. (ii) The supply levels of OR and CS services consistently fall in the middle range among the six services. The supply level of CS ranks high only in Bundle 4. This is because Bundle 4 has the highest proportion of forest area, which is the main supporting land type for CS services, making CS a dominant service within Bundle 4. (iii) The supply level of SC services is consistently the lowest among the six services. SC service levels are relatively high only in Bundle 2, with other ESBs exhibiting lower levels. This is primarily because high SC values are found in high-altitude areas, which align with the hilly terrain of Fujian Province. Bundle 2, located in the Wuyi Mountains, has a higher overall elevation, thus a higher SC service level; other ESBs, despite having SC high-value areas, have a larger proportion of low-elevation regions, resulting in lower overall SC levels. (iv) The supply levels of GP services are highly correlated with the spatial distribution characteristics of arable land within each bundle (Figure 5). GP levels are very low in bundles 1 and 2, where arable land proportion is minimal. The highest GP supply level is observed in Bundle 3, which has the largest proportion of arable land (the only supportive land type for GP).

In these five ESBs, bundle 4 ranks highest in overall service supply levels among the five ESBs, while bundle 3 ranks second, demonstrating a more balanced supply of services. Bundle 5 exhibits the lowest supply levels for both individual services and overall service provision. Therefore, bundles 3 and 4 are considered ideal for increasing occurrence likelihood, whereas bundle 5 is identified as a bundle to be avoided, with efforts needed to reduce its occurrence likelihood.

3.3. Dominant Driving Factors of Ecosystem Service Bundles

Table 2. The relative importance of factors influencing ecosystem service bundles.

Driving Force	Specific Factors	Rank	Relative Importance	Relative Importance of Each Driving Force
Human activities	GDP density	5	6.75	30.9
	population density	3	6.96
	distance from construction land	2	7.07
	distance from highway	16	4.84
	distance from railway	11	5.28
Land management	proportion of cultivated land	13	5.17	28.77
	proportion of woodland	14	5.14
	proportion of grassland	12	5.26
	proportion of water bodies	1	7.48
	proportion of construction land	10	5.73
Climate	annual mean temperature	8	5.95	19.01
	annual total precipitation	6	6.25
	annual solar radiation	4	6.81
Topography	altitude	15	5.02	10.83
	slope	9	5.91
Habitat characteristics	NDVI	17	4.40	10.39
	distance from water bodies	7	5.99

details the relative importance and ranking of 17 factors influencing ESBs. Higher relative importance values and higher rankings reflect a greater impact of these factors on the ESBs. Climate factors generally rank higher in relative importance, while human activity factors show the greatest overall significance. Combined, climate and human activity factors account for 49.91% of the total relative importance, underscoring their substantial influence on ESB formation. Among the five land use type factors, only the proportion of water bodies significantly impacts ESBs, with the highest relative importance among the 17 factors. The proportion of water bodies and distance to construction land are the only two factors exceeding a 7% relative importance threshold, indicating their major role in ES cluster formation. Distance to construction land reflects the extent of human disturbance to ecosystems, further emphasizing the crucial impact of human activity. Additionally, factors such as population density, annual solar radiation, GDP density, and total annual precipitation are also in the top third, which means they are also significant drivers of ESB formation.

3.4. Impact of Dominant Factors on Ecosystem Service Bundles and Optimal Thresholds for Maximum Ideal Bundle Likelihood

Minimizing the occurrence likelihood of undesirable ESBs and promoting the occurrence likelihood of ideal ESBs is beneficial for integrated management of multiple ESs, maximizing overall ecological benefits, and maintaining ecosystem stability. When the occurrence likelihood of undesirable ESBs (bundle 5) is significantly lower than that of ideal ESBs (bundles 3 and 4), the corresponding range of driving factor values indicates the optimal threshold range for those factors.

As illustrated in (Figure 6a), the likelihood of bundle 1 (low GP supply) decreases with a higher proportion of water bodies, suggesting that increased water bodies coverage helps meet crop water needs and thereby enhances the grain production capacity per unit area of farmland. Similarly, the likelihood of bundle 2 (high SC supply) also decreases significantly with an increasing proportion of water bodies. This is because areas with more water bodies usually have lower elevations and more gentle terrain, which reduces the risk of soil loss to some extent [4]. The increasing the proportion of water bodies leads to a significant rise in the likelihood of bundle 4, while the likelihood of bundle 3 also shows an upward trend. This indicates that a higher proportion of water bodies positively affects both the supply levels and the balance of multiple services. However, since the expansion of water bodies is not unlimited, it is essential to identify the range of water bodies proportions that maximizes the occurrence likelihood of bundles 3 and 4 while minimizing the likelihood of bundle 5. When the proportion of water bodies is between 3.7% and 11.4%, the likelihood of bundle 5 is the lowest among the five ESBs, and bundle 4 shows a significant increase. This range is therefore optimal for achieving a high supply clustering of multiple ESs while preserving ecological diversity.

Figure 6b shows that as the distance from construction land increases, the likelihood of bundle 1 (low GP supply) fluctuates upward. In Fujian Province, cultivated land is interspersed with construction land and is generally situated in lower elevations. An increased distance from construction land usually corresponds to higher elevations, a reduced proportion of cultivated land, and a greater proportion of other ecological land types, which can support more ESs except GP [1], explaining the increased likelihood of bundle 1. The likelihood of Bundle 3 increases progressively with the distance from construction land (Figure 6b), suggesting that greater distances from construction land contribute to a more balanced supply of ESs within the ESBs. Conversely, the likelihood of bundle 5 decreases significantly as the distance from construction land increases, indicating that greater distances are associated with higher overall ES supply levels. When the distance from construction land exceeds 7.6 km, the likelihood of bundle 5 is notably lower than that of the other four ESBs, which is more favorable for achieving a high supply and balanced clustering of multiple ESs in the study area.

All ESBs are sensitive to climate change, with their likelihood curves showing significant fluctuations with climate variations (Figure 6c,d). As solar radiation increases, the likelihood of Bundle 4 decreases, while the likelihood of Bundle 5 increases. This may be due to Fujian Province’s subtropical climate, which experiences high sunlight and temperatures year-round, potentially accelerating leaf water evaporation and impacting plant photosynthesis [47], thereby affecting ES supply. When monthly average solar radiation ranges between 346 MJ/m² and 368 MJ/m², the likelihood of bundle 4 is the highest among the five ESBs, with bundle 3 also at a higher level, while bundle 5 has a significantly lower likelihood compared to the other four ESBs (Figure 6c). Similarly, when annual total precipitation falls within the range of 806 mm to 870 mm, the likelihood of bundle 5 is the lowest among the five ESBs (Figure 6d), which is most conducive to a high supply and the balanced clustering of ESs.

As GDP density rises, the likelihood of bundle 1 decreases slightly before stabilizing. This trend occurs because high GDP density areas, predominantly in southeastern Fujian, have abundant farmland that provides more GP. The likelihood curve for bundle 2 (high SC supply) increases slightly before stabilizing, as high-GDP-density areas typically have lower areas of woodland and grassland, which enhance soil loss risk. As population density and GDP density increase, bundle 5 shows a significant rise, indicating that higher levels of human activity negatively impact the supply of ESs. When population density is below 1000 people/km² and GDP density is below 85 million CNY/km², the likelihood of bundle 5 remains relatively low. However, once these thresholds are exceeded, the likelihood of bundle 5 increases significantly.

4. Discussion

4.1. Implications for Land Management

Research has shown that climate and human activities have a significant impact on the formation of ESBs, consistent with the findings of [7,48,49]. The proportion of water bodies, annual total precipitation, and distance to water bodies rank highly in terms of their relative importance for the ESBs formation, indicating that water resources have a major influence on the supply of multiple ESs. Although there are variations in the ranking of different drivers, the differences in their relative importance are minimal (Table 2), suggesting that the formation of ESBs is a result of the combined effects of climate, human activities, land use, topography, and habitat characteristics. Aside from the proportion of water bodies, other land management factors (e.g., proportion of woodland and grassland) rank lower in relative importance. Two possible reasons for this are: (i) The high overall vegetation coverage in Fujian Province results in minimal regional differences in woodland and grassland proportions [14], which may reduce the impact of woodland and grassland area on ESBs. (ii) In studies conducted at large spatial scales, the impact of land use on ESBs may be underestimated when considering the influence of climate, as land use at such scales is predominantly driven by climate [50]. Nonetheless, the impact of land management on ESBs remains an important aspect that cannot be overlooked. Numerous studies have reported that land use changes have a significant influence on the supply of ESs and the spatial heterogeneity of ESBs [21,51,52]. Therefore, based on the influence mechanism of land management factors on ESBs (Figure 7), regulating land use type areas can promote a high and balanced supply of multiple ESs.

When the proportion of cultivated land falls within the range of 16.5% to 36%, the likelihood of occurrence of bundle 5 is lower than that of bundles 3 and 4, with bundle 4 exhibiting the highest likelihood of occurrence (Figure 7a). This range is most favorable for enhancing and balancing the supply levels of six ESs in the study area. When the proportion of cultivated land does not exceed 59%, the probabilities of occurrence for bundles 3 and 4 are generally higher, indicating better overall supply levels and balance of the six ESs in the study area. However, as the proportion of cultivated land continues to increase, the probabilities of occurrence for bundles 3 and 4 significantly decrease, while the likelihood of occurrence for bundle 5 increases (Figure 7a). Due to the limited support of cultivated land for all services except for GP, a high proportion of cultivated land may exacerbate competitive exclusion among services [19], leading to trade-offs between services. Therefore, the proportion of cultivated land should be maintained within the range of 16.5% to 36%, and should not exceed 59%.

With the increase in the proportion of woodland, the likelihood curve for bundle 3 exhibits a bimodal distribution (Figure 7b). The first peak occurs when the proportion of woodland is 6.5%, a situation primarily observed in construction land where the limited number of ecological land types results in low levels of multiple services simultaneously. The second peak (the highest likelihood) occurs at a woodland proportion of 62%, indicating that a 62% woodland proportion is most conducive to the balanced development of ESs in the study area. When the proportion of woodland is between 31.7% and 39.6%, the likelihood of occurrence for bundle 4 is at its highest, reflecting the strongest overall supply level of the six ESs in the study area. However, this corresponds to a trough in the likelihood curve for bundle 3. This situation predominantly occurs in forest-grassland transition zones, where forest and grassland cannot support GP but provide strong support for other services [9]. Thus, despite GP being at a disadvantage, the overall supply capability of the six services is the strongest. When the proportion of woodland exceeds 54.3%, the likelihood of occurrence for bundle 5 is lower than for bundles 3 and 4, suggesting that this proportion is most favorable for enhancing the overall output of ESs in the study area and promoting synergy between services.

The increase in the proportion of water areas promotes the enhancement and balance of multiple ecosystem service supply levels (the occurrence probabilities of bundles 3 and 4 show a growing trend) (Figure 6a), suggesting that an appropriate increase in water area is beneficial. When the water area proportion is between 3.7% and 11.4%, it can effectively prevent multiple ESBs from exhibiting low supply clustering states. When the grassland proportion is between 23.5% and 50%, the occurrence likelihood of bundle 5 is significantly lower than that of the other four ESBs, while the occurrence probabilities of bundles 3 and 4 remain relatively high (Figure 7c). This scenario is most conducive to improving and balancing the supply levels of six ecosystem services in the study area. As the proportion of grassland continues to increase, the occurrence likelihood of bundle 5 gradually rises, surpassing that of bundle 4 when the grassland proportion reaches 65.5%. Therefore, it is advisable not to exceed a grassland proportion of 65.5%.

Human activity intensity significantly reduces the supply levels of multiple ESs: the closer the distance to construction land, and the higher the proportions of built-up land, population density, and GDP density, the higher the likelihood of occurrence of bundle 5 (Figure 6b,e,f and Figure 7d). Variables such as population density, GDP density, distance to built-up land, and proportion of built-up land represent the rural-urban gradient and have been shown to be important factors in land use change and resource utilization [21,53,54], typically indicating the level of urbanization. Urbanization consumes ecological land types and exacerbates trade-offs between ESs. As urbanization intensity increases, the synergy in the supply capacity of multiple ESs diminishes. Consistent with previous studies [55], urban areas have lower ES supply, and rapid expansion of built-up areas increases the extent of cold spots [7]. Therefore, expansion of built-up land should be strictly controlled. When the proportion of built-up land does not exceed 10.1%, the probabilities of occurrence for bundles 3 and 4 are higher, while the likelihood for Bundle 5 is the lowest among the five ESBs, which is optimal for enhancing the overall output of ESs and promoting synergy between services in the study area.

4.2. Comparison with Results from Other Regions

Mei et al. [56] analyzed the spatiotemporal evolution patterns of ecosystem service bundles in Anhui Province, China, at a 30 m × 30 m grid scale. Their results indicated that among socioeconomic factors, built-up land exerted the most significant influence, which is consistent with the conclusions of this study. Most existing research has focused on the driving factors of ecosystem service supply levels, which in turn affect the formation of ecosystem service bundles. At a 1000 m × 1000 m grid scale, Li et al. [49] using the Beijing–Tianjin–Hebei urban agglomeration as the study area, pointed out that water resources are the key factor influencing ecosystem service supply, a finding that aligns with our results. Similarly, Wang et al. [57] at the 1000 m × 1000 m grid scale in the Zhangjiakou–Chengde region, found that once land use and development intensity exceed critical thresholds, ecosystem service values decline significantly. This further corroborates the viewpoint of this study: ecosystems exhibit multi-stability, and identifying the optimal threshold ranges of driving factors is essential for the scientific management of ecological resources.

4.3. Model Validation

To ensure the accuracy of the MC-EBM model results and to mitigate optimistic estimates that may arise from biased classifiers on imbalanced datasets, this study also evaluated the MLR model and compared its accuracy rate and Balanced Accuracy (BACC) with those of the MC-EBM model. Higher values of accuracy rate and BACC indicate a better model fit [24]. As shown in

Table 3. Model Performance Comparison.

Model Accuracy Parameter (%)	MLR	MC-EBM
Accuracy rate for bundle 1	11.7	83.9
Accuracy rate for bundle 2	45.1	90.1
Accuracy rate for bundle 3	26.0	90.0
Accuracy rate for bundle 4	32.5	80.8
Accuracy rate for bundle 5	68.4	86.9
BACC metric	37.1	84.5

Note: MLR is multivariate logistic regression, and MC-EBM is multiclass explainable boosting machine.

, the accuracy of the MC-EBM model exceeds that of the MLR model. Furthermore, the MLR model, being linear, depends on prior assumptions (linear regression equations) to model the relationship between variables, which may lead to biases relative to the actual conditions. In contrast, the MC-EBM model can capture the true correlations in the ecological processes by learning and training on the original dataset. Once integrated into an API-interpretable algorithm, the MC-EBM model allows for visualization of the variable correlation processes. The response curves of ESBs to various factors are nonlinear and exhibit periodic fluctuations (Figure 6 and Figure 7), which cannot be accurately represented by a single linear equation.

5. Conclusions

This study introduces an interpretable multi-class machine learning model (MCEBM-API) and, from a nonlinear and threshold perspective, proposes a methodological framework for exploring the driving mechanisms of ESBs. The study addresses the three questions raised in the introduction:

(i)

Water resources, climate, and human activities significantly influence ESB formation. Of the 17 “natural-socioeconomic” factors analyzed, the proportion of water bodies, distance from construction land, annual solar radiation, total annual precipitation, population density, and GDP density are the primary determinants affecting ES cluster formation in Fujian Province.

(ii)

An increased proportion of water bodies enhances and balances the supply levels of multiple ESs, while higher human activity intensity markedly reduces the supply levels of various ESs. Additionally, all ES clusters exhibit high sensitivity to climate change.

(iii)

At the 1 km × 1 km grid scale, each driving factor has an optimal threshold range that increases the likelihood of forming ideal ESBs: cultivated land should be maintained between 16.5% and 36% (and not exceed 59%), while woodland should account for more than 54.3%, water bodies for 3.7–11.4%, grassland for 23.5–50%, and built-up land should not exceed 10.1%. Regulating land use type areas within these ranges can foster the formation of ESBs with high and balanced supply levels, while preventing the emergence of low-supply ESBs.

(iv)

In exploring ESB formation mechanisms, MCEBM-API significantly outperforms multinomial logistic regression and demonstrates strong generalizability.

These findings deepen our understanding of ES spatial clustering patterns and provide support for the scientific management of ecosystem services. Nonetheless, this study has limitations: it only used data from a single year and examined ESB formation at the 1 km × 1 km grid scale, while mechanisms may differ across spatial scales. The proposed methodological framework is broadly applicable and could be further extended to studies in different regions, at multiple spatial scales, and with multi-temporal datasets to validate its robustness and applicability.