Optimizing Spatial Pattern of Water Conservation Services Using Multi-Scenario Land Use/Cover Simulation and Bayesian Network in China’s Saihanba Region

Abstract

Optimizing the spatial pattern of water conservation services (WCSs) is essential for enhancing regional water retention and promoting sustainable water resource management. The Saihanba region, a critical ecological barrier in northern China, has experienced severe degradation due to historical over-logging, leading to weakened WCS functions. This study used remote sensing techniques to interpret land use/land cover change (LULC) and combined it with meteorological and basic ecological data to assess changes in WCS capacity in the Saihanba region, China, under multiple 2035 scenarios using CA-Markov and Bayesian network models. The Bayesian belief network identified priority areas for spatial optimization. Results showed the following: (1) The spatial distribution patterns of WCSs showed a strong dependence on land-use types, with both forest and grassland areas demonstrating superior water conservation capacity compared to other land cover categories; (2) although total WCS capacity varied across scenarios, spatial distribution remained consistent—high-value zones were mainly in the south and central-east, while lower values occurred in the west; and (3) WCS areas were categorized into key optimization, ecological protection, and general management zones. Notably, the Sandaohekou Forest Farm and the western Qiancengban Forest Farm emerged as critical areas requiring urgent optimization. These findings offer practical guidance for spatial planning, ecological protection, and water resource governance, supporting long-term WCS sustainability in the region. The study also contributes to cleaner production strategies by aligning ecosystem service management with sustainable development goals.

1. Introduction

Ecosystem services (ESs) refer to all the benefits humans derive from ecosystems, including provisioning, regulating, cultural, and supporting services [1]. Regulating services, central to ES, encompass functions such as climate regulation, hydrological regulation, and soil erosion control [2,3]. Water conservation services (WCSs), a key component of regulating services [4,5,6], refers to the ability of ecosystems to intercept and temporarily retain precipitation, thereby regulating surface runoff, enhancing available water resources, and preventing soil erosion. Land use is a major human-driven change that profoundly impacts ecosystem hydrology [7]. The responses of hydrological processes to changes in land use are rapidly altering WCS functions [8]. In the foreseeable future, driven by national policies, social demands, and environmental changes, land-use types and structures are expected to undergo significant transformations, posing challenges to the sustainable development of regional water resources [7,8].Therefore, given the complexity of land-use changes, simulating WCS dynamics under future scenarios and clarifying its response mechanisms are crucial for national land planning, water conservation, and regional sustainable development.

Extensive research has been conducted on the impacts of land use/cover change (LULC) on WCS [8,9,10]. Strategic adjustments in land-use patterns are considered effective in enhancing WCS capacity. These adjustments directly affect the underlying surface construction, serving as a primary manifestation of human-induced environmental transformations and a key driving factor of WCS alterations [9]. Researchers have increasingly employed various land-use forecasting models integrated with methods such as the water balance model to assess future changes in WCS capacity [6,11]. For example, Zeng et al. [12] applied the water balance method to quantify WCS capacity by incorporating precipitation, evapotranspiration, and surface runoff, providing a basis for Bayesian network analysis in the Weihe River Basin, China. With methodological advancements, increasingly sophisticated land-use simulation models have developed, evolving from early quantitative forecasting methods like Markov and logistic regression models to integrated approaches such as the CA-Markov model, which effectively captures spatial dynamics of land-use changes [13,14,15].

Despite quantitative analyses of future WCS, spatial optimization remains underexplored. The Bayesian Belief Network (BBN) offers a promising solution to this challenge. As a powerful tool for managing uncertainty, BBN represents relationships among multiple variables through conditional probabilities [16,17]. It excels in learning and reasoning under limited and uncertain information, rendering it well-suited for ES modeling. As a probabilistic graphical network, BBN identifies the key factors influencing WCS capacity, quantifies their impact, and visualizes research data, effectively optimizing spatial patterns shaped by multiple factors [18,19,20].

Recent studies have underscored the significant impact of LULC on WCS through alterations in surface cover, hydrological processes, and interactions with climate change and human activities. For instance, Zeng et al. [21] demonstrated that WCS responds strongly to changes in precipitation and temperature, with notable differences under varying environmental conditions. Nsiah et al. [22] showed that changes in land cover significantly alter surface runoff, evapotranspiration, and water availability in tropical regions. Yang et al. [23] also emphasized the role of vegetation greening in enhancing WCS through its influence on evapotranspiration and vegetation changes. These studies highlight the need to consider multi-factor interactions to better understand and predict the dynamics of WCS.

Saihanba (SHB), historically part of the Qing Dynasty’s Mulan Weichang hunting ground, underwent severe degradation due to excessive logging, transforming into a sandy wasteland by the 1950s and becoming a major source of sandstorms affecting Beijing [24]. In response, the Saihanba Mechanical Forest Farm was established in 1962 and has since been managed through mechanized afforestation and reforestation efforts, transforming the region from degraded sandy land into the world’s largest plantation forest. Integrated restoration strategies have driven key ecological projects, significantly enhancing ecosystem quality and establishing an ecological security barrier [1,25,26]. Over the past two decades, the adoption of multi-objective management approaches has further accelerated ecosystem recovery, earning SHB the UNEP Champions of the Earth award in 2017 [1,25].

Despite significant ecological achievements, SHB still faces challenges, as prolonged high-density forestry operations may lead to soil degradation and weaken the WCS functions, posing risks to long-term ecological sustainability and resource security [25,26,27]. The present study focuses on the unique characteristics of the SHB region, adhering to the principles of ecological civilization and the delineation of key functional zones. A scenario simulation approach was employed to predict land-use scenarios for the SHB region in 2035. Our research offers multi-dimensional insights into land use in the SHB region and develops a BBN for the WCS optimization in land-use planning. It also provides a theoretical framework for high-quality, stable, and sustainable development through environmentally responsible land-use strategies.

2. Data and Methods

2.1. Study Region

The study region, SHB, is located in Weichang County, Hebei Province, China (116°53′–117°39′ E, 42°4′–42°36′ N), at the southern edge of the Horqin Sandy Land in Inner Mongolia [25]. The region covers an area of approximately 933.33 km² and encompasses six sub-forestry stations, with an elevation range from 1010 to 1939 m (Figure 1). The region has a temperate continental monsoon climate, with semi-arid and semi-humid conditions. The average annual temperature is −2.7 °C, and precipitation occurs mainly from June to August, averaging 460.3 mm annually. Snow cover lasts for seven months, and the frost-free period is only 64 days each year [1,25]. The selected study years (2002, 2012, and 2022) correspond to pivotal stages in SHB’s ecological restoration: the designation of the region as a National Nature Reserve in 2002, the transition to multifunctional forest management in 2012, and the culmination of decades-long restoration efforts in 2022.

2.2. Data Sources and Preprocessing

The dataset used in this study comprises seven major categories: geographic data, remote-sensing imagery, climate, soil, hydrology, vegetation, and topography. Geographic data include administrative boundaries, rivers, and road vectors. Remote-sensing imagery provides land-use/land-cover maps for three time points. Climate data include annual precipitation, mean annual temperature, and evapotranspiration. Soil data consist of soil type and thickness. Hydrological data include surface runoff, water yield, and water retention. Vegetation data comprise NDVI, vegetation type, and vegetation coverage. Topographic data consist of DEM and slope. The dataset spans the period from 2002 to 2022. All raster layers were resampled to a consistent spatial resolution of 30 m resolution and co-registered across timeintervals to ensure spatiotemporal consistency in model inputs (see Supplementary Tables S1 and S2 for detailed information) [28,29].

2.3. Research Framework

This study employed a three-step methodological framework: (a) three distinct scenarios—economic development, natural development, and ecological protection—were developed to simulate and project LULC changes in the region by 2035. These scenarios were formulated based on historical land-use trends and relevant regional policies. (b) Using the simulated 2035 LULC data, spatiotemporal variations in WCS capacity were assessed by integrating the water balance method with the InVEST model, supported by ArcGIS and Netica software. (c) A Bayesian Belief Network (BBN) was constructed to explore the complex interactions among multiple influencing factors. Sensitivity analysis was conducted to identify the most influential variables, and model validation was carried out to ensure accuracy. To account for uncertainty, variations in LULC and WCS capacity were incorporated into the BBN framework. This enabled the delineation of WCS spatial distribution patterns under the three scenarios. Based on the critical thresholds of key variables, priority areas for WCS optimization were identified. Finally, tailored spatial optimization strategies for the SHB region were proposed (Figure 2).

2.4. Multi-Scenario LULC Simulation Based on the CA-Markov Model

The CA-Markov model simulated the land-use patterns under three scenarios outlined in Section 2.3 through five steps.

(a)

The transition areas between each land-use type in 2002, 2012, and 2022 were used as elements in the Markov state transition probability matrix, which illustrates the spatiotemporal dynamics of land use, as expressed by the following formula:

$$U_{ij}=\left[\begin{array}{ccc}{U}_{11}& \dots & U_{1n}\\ ⋮& ⋮& ⋮\\ U_{n1}& \dots & U_{nn}\end{array}\right]$$

(1)

where U_ij represents the transition of land-use type from i to j during the study period, and n denotes the number of land-use types.

Based on the characteristics of land-use types (e.g., cropland, forest land, grassland, etc.), driving factors were selected as follows: slope, elevation, distance to roads, and distance to water bodies. These factors were used to calibrate the CA-Markov model parameters for different scenarios. The calibration process involved establishing a transition probability matrix and suitability map sets. In accordance with land-type conversion characteristics and relevant policy plans, restrictive and suitability factors were selected to calculate the potential distribution of conversion types. Additionally, suitability distribution maps for each land-use type were developed, and tailored suitability maps were created for each scenario.

(b)

Using 2002 and 2012 as baseline years, land-use patterns in 2022 were predicted and validated against actual data. Constrained by SHB’s planning framework, suitability factors were weighted using the entropy method [25]. Validation yielded an accuracy of 92.41% and a Kappa coefficient of 0.851, indicating an “almost perfect” agreement [30]. These results were consistent with previous studies [5,25], confirming the model’s reliability for future land-use projections.

(c)

In the natural development (ND) scenario, future land-use changes were projected based on the region’s natural development trends. The ecological protection (EP) scenario prioritized location-specific ecological improvements and incorporated relevant restrictions, aligning with the “Master Plan for Saihanba Mechanical Forest Farm in Hebei Province (2017–2030)”. The economic development (ED) scenario integrated ecological protection with economic growth by fostering green industries, establishing international research bases, and expanding tourism services, following the “Ecological Protection Plan for Forest and Grassland in Saihanba Mechanical Forest Farm and Surrounding Areas (2020–2035)”.

(d)

Based on land-use maps for 2012 and 2022, the spatial allocation of key indicators under the three scenarios was achieved using interpolation methods, generating potential distribution maps for land-use type conversions. IDRISI software (ver. 18.0; Clark Labs, Worcester, MA, USA) was then employed to simulate land-use changes and distribution patterns for 2035. The CA-Markov model was applied to predict both the quantity and spatial distribution of land-use changes based on transition probabilities and spatial rules derived from historical data.

2.5. Quantification and Validation of WCS Capacity

The water balance method [28] was applied to estimate the SHB region’s WCS capacity in 2022. The WCS capacity for different land-use types was then calculated and summed to obtain the total WCS capacity of the study region, as calculated using the following formula [5]:

$$Q=10\times \sum S_i(P_i-E_i-R_i)$$

(2)

where Q is the total WCS capacity of the study region (m³/a), S_i is the area of land-use type i (hm²), P_i is the annual precipitation for land-use type i (mm/a), E_i is the annual evapotranspiration for land-use type i (mm/a), and R_i is the surface runoff for land-use type i (mm/a).

To further validate the calculation results of WCS capacity, the Annual Water Yield module in the InVEST model [29] was employed for additional calculations, with the formula as follows:

$$Y_{\left(x\right)}=(1-{\displaystyle \frac{AE{T}_{\left(x\right)}}{P_{\left(x\right)}}})\times P_{\left(x\right)}$$

(3)

where Y(x) is the annual water yield of grid cell x (mm), AET(x) is the annual actual evapotranspiration of grid cell x (mm), and P(x) is the annual precipitation of grid cell x (mm).

To verify the reliability of the quantitative results, a method integrating the topographic index, soil saturated hydraulic conductivity, and velocity coefficient was used to fit the WCS capacity, with the following formula [5]:

$$C_x=min(1,{\displaystyle \frac{249}{V_{lx}}})\times min(1,0.3\times T_x)\times min(1,{\displaystyle \frac{K_{sat,x}}{300}})\times Y_x$$

(4)

$$T_x=\mathit{lg}({\displaystyle \frac{A_{drainage,x}}{d_{soil,x}\cdot s_{percent,x}}})$$

(5)

where C_x is the water conservation capacity (mm), V_lx is the velocity coefficient of land-use type l_x, T_x is the topographic index, K_sat,x is the soil saturated hydraulic conductivity (mm/d), Y_x is the water yield calculated by the InVEST model (mm), A_drainage,x is the drainage area of the region, d_soil,x is the soil thickness (mm), and S_percent,x is the slope percentage.

The coefficient of variation for both methods was less than 5%, and a significant positive correlation was observed between the two methods (y = 1.608x − 305.822; p < 0.01; R² = 0.753), confirming the accuracy of the results.

2.6. Spatial Pattern Optimization of WCS Capacity Based on BBN

The BBN for WCS capacity was constructed using Netica software (NORSYS Software Co., NORSYS, Québec, Canada). Each node in the BBN is characterized by a conditional probability table (CPT), the discrete states of the node variables, and their corresponding probability distributions [16,17]. The CPT quantifies the relationship between a parent node X and a child node Y, with each row of the table representing the state combinations of the parent node and the conditional probability P(y/x). The probability distribution of the child node is determined by the probability distribution of the parent node and the CPT. In the absence of parent nodes, the probability distribution is the prior probability distribution P(x). For example, the CPT for the child node of evapotranspiration is presented in Supplementary Table S3. The joint probability of all variables in the BBN was calculated by multiplying the conditional probability distributions of all nodes, as shown in the following formula [16]:

$$P(X_1,X_2,\dots X_n)=\prod _{i=1}^{n}P(X_i|parents\left(X_i\right))$$

(6)

where P (Xi|parents(Xi) represents the probability distribution of the variable node X_i, and X₁, X₂, …, X_n are the variable nodes in the network.

Considering data availability, underlying WCS mechanisms, regional conditions, and relevant literature [12,31,32], 13 WCS-related factors were selected as node elements for the BBN (measurement methods detailed in Section 2.2). Among these factors, slope and elevation reflect the influence of topography on ES [33,34]; NDVI, vegetation type, vegetation cover, soil thickness, and soil type comprehensively represent the impact of vegetation on ES, reflecting vegetation growth and coverage as well as soil conditions [35,36,37]. WCS is mainly controlled by precipitation, evapotranspiration, water yield, LULC, and surface runoff [22]. Evapotranspiration is mainly influenced by temperature and vegetation type [23]. Surface runoff is primarily associated with precipitation, slope, land use, and soil type [23,27].

Data discretization was performed to generate input data for BBN construction using two methods: the natural breakpoint method, which minimizes within-group variance to identify natural thresholds, and the custom interval method, which classifies data based on predefined ranges informed by expert knowledge and prior research (see Supplementary Table S4 for classification criteria). An uncertainty propagation framework [20] was applied to perform sensitivity analysis of BBN input variables, thereby assessing model robustness. Netica software was then used to construct the WCS BBN. To accurately determine relationships between driving factors and target variables, parameter learning was conducted to validate the BBN model [16,21,38]. CPTs for each network node were calculated using the following formula [39]:

$$P(B│A)={\displaystyle \frac{P\left(AB\right)}{P\left(A\right)}}$$

(7)

where P(AB) denotes the joint probability of events A and B, representing the probability of both events occurring simultaneously. P(A) and P(B) are the prior probability of events A and B, respectively.

To further validate the BBN model’s predictive accuracy, 10,000 data points were randomly selected and combined with the remaining data to form training and validation sets. The training set was input into the BBN model to generate CPTs. Model accuracy was assessed using a confusion matrix, calculated as follows [12]:

$$Error={\displaystyle \frac{\sum Casef}{\sum Casef+\sum Caset}}\times 100\%$$

(8)

where $\sum Casef$ and $\sum Caset$ represent the number of data points in which the predicted values of vulnerable nodes differ from and are consistent with the actual values of WCS capacity, respectively.

After constructing the BBN for WCS, key variables and their state subsets were identified through CPT and sensitivity analysis. The subset space was then visualized to show zones where key variables and states of WCS nodes co-occur, thereby delineating optimization zones. This process involved three steps.

(a)

Sensitivity analysis was used to evaluate how the WCS node responds to changes in other influencing factors. The entropy reduction value quantifies the impact of each input variable on the target variable, with the following formula [39]:

$$ERV=V\left(ES\right)-V\left(ES|I\right)=\sum _{s}p(s)\times {\left(s-E\left[ES\right]\right)}^2-\sum _{s}p(s|I)\times {\left(s-E\left[ES\right|I]\right)}^2$$

(9)

where $V\left(ES\right)$ and $V\left(ES|I\right)$ represent the entropy of the WCS without and with considering the input node I, respectively. s is the state of the output variable, $p\left(s\right)$ is its probability, and $p\left(s\right|I)$ is the conditional probability for I. $E\left[ES\right]$ denotes the expected WCS value, while $E\left[ES\right|I]$ represents the expected WCS value given the state of the input node I.

(b)

Based on the BBN model, the conditional probabilities of influencing factor pairs were calculated. The states with the highest conditional probabilities, corresponding to different WCS capacity levels, were selected as the critical states of the variables. The WCS capacity was categorized into five levels: highest, high, medium, low, and lowest (designated as Subset I, II, III, IV, and V, respectively).

(c)

After identifying key state subsets for WCS under three scenarios, the one with the largest Subset V area was selected as optimal. Within this scenario, Subset I was prioritized for improvement. Optimization zones were then defined based on this subset and the latest land-use policies.

3. Results

3.1. Land Use/Cover Changes from 2002 to 2035

The distribution of land-use types in the SHB region exhibited significant spatial heterogeneity from 2002 to 2035. Forest and grassland were the predominant land-use types, accounting for over 90% of the total area of the SHB region, while other land-use types covered less than 10% (Figure 3). Forests, exhibiting extensive and dense distribution, covered 78.77% to 90.40% of the SHB region (

Table 1. Proportion of land-use changes in the Saihanba region from 2002 to 2035 (%).

LULC	2002	2012	2022	2035 Natural Scenario	2035 Conservation Scenario	2035 Development Scenario
Cropland	1.41	1.44	0.91	1.20	0.91	1.49
Forest	81.93	78.77	90.23	87.4	87.3	87.84
Grassland	2.19	13.99	3.94	6.35	6.73	5.58
Wetland	3.79	4.53	3.29	3.51	4.39	2.97
Urban land	0.27	0.19	0.70	0.69	0.81	1.81
Unused land	10.41	1.08	0.93	0.85	0.34	0.31

). Grasslands, in contrast, were highly fragmented and dispersed within forested areas (Table 1). Wetlands were predominantly distributed in a linear pattern across the northwestern part of the SHB region, while cropland and built-up land, limited in extent, were mainly located in the western and central-eastern parts (Table 1).

During the period from 2002 to 2035 (Table 1), land-use types in the study region showed distinct changing trends. Specifically, the proportions of cropland and wetlands initially increased, then decreased, before rising again. In contrast, the proportion of grassland initially increased, followed by a decline. The proportion of forest land decreased from 81.93% in 2002 to 78.77% in 2012, before steadily rising to 90.40% by 2035. Built-up land steadily increased throughout the study period, while unused land continuously decreased.

Furthermore, the alteration in the spatial distribution of land-use types from 2022 to the three simulated scenarios for 2035 was also evaluated (Figure 3). Compared to 2022, under the 2035-EP scenario (Figure 3a), the areas of cropland and forest land remained stable, while wetland and built-up land increased, and the unused land decreased significantly by 62.37%. Under the 2035-ND scenario (Figure 3b), the areas of forest land, wetlands, and unused land decreased slightly, while the areas of cropland, grassland, and built-up land increased by 31.87%, 7.14%, and 68.78%, respectively. Under the 2035-ED scenario (Figure 3c), the areas of grassland, cropland, and built-up land increased by 41.62%, 63.74%, and 157.14%, respectively, while the areas of forest land, wetlands, and unused land all decrease to some extent. These findings suggest that the EP scenario performed optimally in terms of ecological protection and sustainability, while the ED scenario exerted the most significant influence on natural ecosystems.

During the period from 2002 to 2012 (Figure 4), 78.33% of unused land, 6.03% of cropland, and 7.46% of wetlands were converted to grassland. Additionally, 12.32% of unused land and 8.46% of wetlands were transformed into forest. Furthermore, 86.15% of grassland and 75.65% of unused land were converted to forest. By 2035 (Figure 4), the proportion of unused land was the lowest in the study region, suggesting optimized land use. The magnitude of land-use transitions between 2002 and 2012, as well as between 2012 and 2022, was greater than that observed in the scenarios for 2022 and the simulated results for 2035, consistent with the overall developmental trend of the SHB region (Figure 4).

3.2. Spatiotemporal Analysis of WCS Capacity

Over the three periods (i.e., 2002–2012, 2012–2022, and 2022–2035), the overall trend in WCS capacity of the SHB region was characterized by an initial decrease, followed by a subsequent increase. High levels of WCS capacity were observed in the southern and central-eastern parts of the SHB region, while low levels were found in the western part (Figure 5a–c).

In terms of specific scenarios simulated for 2035 (Figure 5e,f), the EP scenario showed significant improvements in WCS capacity in the central, southern, and eastern parts of the SHB region compared to 2022 (Figure 5d). The ND scenario exhibited a distribution of WCS capacity largely consistent with that of 2022 (Figure 5e). In contrast, the ED scenario witnessed a slightly decreased WCS capacity in the southern and central parts of the SHB region relative to 2022 (Figure 5f).

During the study period, the average WCS volume in the SHB region was relatively high in 2002, with a depth of 307.35 mm per unit area. By 2012, this value had decreased to 206.83 mm per unit area before slightly rebounding to 228.43 mm per unit area in 2022. According to the simulated results for 2035, under the EP scenario, the WCS volume showed an increasing trend compared to 2022, reaching 329.27 mm per unit area. In the ND scenario, the WCS volume remained close to that of 2022, at 246.44 mm per unit area. However, under the ED scenario, a decreasing trend was observed over the period, with the depth dropping to 211.81 mm per unit area (

Table 2. Average water conservation service capacity values for different land-use types from 2022 to 2035 (mm).

LULC	2002	2012	2022	2035-EP	2035-ND	2035-ED
Cropland	253.60	268.37	265.77	265.56	279.61	287.36
Forest	213.01	294.51	302.39	309.08	303.2	301.21
Grassland	331.46	298.32	310.16	313.40	306.54	279.25
Wetland	434.03	308.68	404.12	485.03	417.74	344.69
Urban land	0.74	0.17	0.58	3.95	0.66	0.72
Unused land	8.82	4.26	2.59	6.11	1.27	2.32

Notes: LULC, land use/land cover. EP, ecological protection scenario. ND, natural development scenario. ED, economic development scenario.

).

Furthermore, the overall WCS volume in the SHB region exhibited a heterogeneous spatial distribution pattern (Figure 5). In 2002 and 2022, the central-southern and eastern parts of the SHB region displayed high WCS volumes, particularly in the Dahuanqi Forest Farm, Sanxiang Forest Farm, and Yinhe Forest Farm (Figure 5a,c). In contrast, poor WCS volume was observed in these parts of the SHB region in 2012, particularly in the Sanxiang Forest Farm (Figure 5b). Moreover, parts with high WCS value were sparsely distributed in 2012 (Figure 5b), whereas in 2002 and 2022, their distribution was relatively concentrated (Figure 5a,c). Under the simulated EP scenario for 2035, WCS volumes in the Sanxiang Forest Farm and Yinhe Forest Farm showed a remarkable increase compared to 2022 (Figure 5d), while changes under the ND and ED scenarios were minimal (Figure 5e,f).

3.3. WCS Capacity of Different Land-Use Types

The quantitative WCS capacity values for different land-use types in the SHB region (Table 2) reveal significant differences, with the overall ranking as follows: wetlands > grasslands > forests > croplands > unused land > built-up land.

From 2002 to 2012, WCS capacity increased in cropland and forests, while an opposite trend was observed in other land-use types. From 2012 to 2022, WCS capacity increased in forests, grasslands, and wetlands, while no changes were observed for the other land-use types. Projections for 2035 suggest that under the EP scenario, WCS capacity would increase across all land-use types except for croplands. Under the ND scenario, increases were expected in croplands and wetlands, while declines were projected for grasslands and unused land, with no significant changes in forests and built-up land. Under the ED scenario, WCS capacity would increase in croplands and built-up land, whereas declines were projected for the other land-use types.

3.4. Optimization of Spatial Patterns of WCS Capacity Under Multiple Scenarios

The prediction results from the BBN model (Figure 6) show that, under current environmental and land-use conditions, the overall WCS capacity in the SHB region was at a relatively low to medium level. The “lowest” and “medium” categories accounted for 43.1% and 37.3% of the total, respectively.

Uncertainty analysis (

Table 3. Sensitivity analysis of water conservation service capacity to each node.

Actual Results of Water Conservation	Mutual Information	Percentage/%	Entropy Reduction	Uncertainty Coefficient
Water conservation	0.89037	100	0.1857876	-
Precipitation	0.09061	10.2	0.0074825	0.123
Land use/cover change	0.05138	5.77	0.0025989	0.087
Surface runoff	0.03417	3.84	0.0049141	0.065
Evapotranspiration	0.01857	2.085	0.0025296	0.042
Water yield	0.01027	1.153	0.0021177	0.038
Slope	0.00425	0.477	0.0007488	0.015
Temperature	0.00044	0.0496	0.0000048	0.008
Soil types	0.00018	0.0206	0.0000019	0.005
Digital elevation model	0.00011	0.01306	0.0000017	0.003
Vegetation cover	0.00011	0.01235	0.0000017	0.003
Vegetation types	0.00003	0.00323	0.0000013	0.001
Normalized difference vegetation index	0.00001	0.00137	0.0000008	0.001
Soil depth	0.00001	0.00124	0.0000002	0.001

) shows that the top five variables contributed to a cumulative uncertainty coefficient of 0.355. Precipitation uncertainty caused a ±12.3% fluctuation in WCS estimates, while land-use classification errors shifted the optimized boundary by <0.5 km. The contributions of other variables remained below 5%, indicating high model robustness. Predictive performance was further validated using a confusion matrix, which showed an overall accuracy of 79.17% (Supplementary Table S5), supporting the model’s reliability in forecasting WCS capacity across multiple scenarios in 2035.

The distribution probabilities of WCS capacity status across the three scenarios were similar (Supplementary Figure S1), with the highest probability of water conservation depth per unit area ranging from 50 mm to 300 mm, while values exceeding 300 mm were less probable. However, variations among the scenarios were observed, with the EP scenario exhibiting the highest probability of higher WCS capacity levels, followed by the ND and ED scenarios (Supplementary Figure S1).

Land-use types, particularly wetlands, contributed the most to WCS capacity, with a rate of 22.48% under the “highest” category, as indicated by the contribution rate analysis of key variables (Supplementary Table S6). Among climatic factors, precipitation was positively correlated with WCS capacity, with the maximal probability of WCS capacity being classified as “lowest” under “lowest” or “low” precipitation conditions. As precipitation increased to “high” or “highest” levels, WCS capacity was more frequently classified as “medium” or above (Supplementary Table S6).

The impact of surface runoff on WCS capacity was nonlinear (Supplementary Table S6). When surface runoff was at the “lowest” status, WCS capacity had a 62.10% probability of being classified as “lowest”. As surface runoff increased to “high” status, the probability of WCS capacity falling into “high” categories peaked at 51.45%. However, at the “highest” runoff level, the probability of WCS capacity being classified as “medium” or above decreased compared to the “high” status. Evapotranspiration initially enhanced WCS capacity but later showed a diminishing effect. Conversely, increased water yield consistently contributed to higher WCS capacity. Moreover, changes in NDVI, topography, and soil factors had minimal impact on the most probable state of WCS capacity, leading to only slight probability fluctuations.

The sensitivity analysis (Table 3) identified precipitation, land-use type, surface runoff, evapotranspiration, and water yield as the five most influential variables, ranked by entropy loss. These variables exhibited high sensitivity to fluctuations in WCS capacity, thus being set as key determinants under different scenarios.

The analysis of key variables and state subsets (Figure 7a–c) demonstrated consistent trends across the simulated EP, ND, and ED scenarios by 2035. Under the EP scenario, the areas covered by Subset I and Subset V were minimal and most extensive, respectively, suggesting an optimal WCS capacity state. The study region was further divided into key optimization, ecological protection, and general management zones based on visualization results from the five key state subsets (Subsets I–V; Figure 7). Subset I, primarily located in the Sandaohekou Forest Farm and the western part of the Qiancengban Forest Farm, corresponded to the lowest WCS capacity and was designated as a key optimization zone. In contrast, Subset V, corresponding to the highest WCS capacity, was mainly distributed in the Sanxiang and Yinhe Forest Farms in the central-eastern SHB region and was therefore designated as an ecological protection zone. Finally, the other transitional areas were designated as a general management zone (Figure 8).

4. Discussion

4.1. The Impact of LULC on WCS Capacity

WCS is essential for sustaining regional water cycles and ecological balance. As a critical ecological barrier in northern China, the SHB region (Figure 1) has recently experienced severe ecological degradation, hindering the full realization of its WCS functions. This study investigated LULC impacts on WCS in the SHB region, indicating that land-use types differed in their contributions (Supplementary Table S6), influencing regional WCS capacity and overall ES performance.

Wetlands effectively intercept and store precipitation, reduce surface runoff, and enhance soil water retention capacity due to their unique ecosystem structure and functions [24]. In the SHB region, wetlands contributed 22.48% to the “highest” WCS category (Supplementary Table S6), underscoring their irreplaceable role in regulating the water cycle and enhancing water storage [24]. Interestingly, grasslands contributed more to WCS capacity than forests (Supplementary Table S6), contrasting with Wang et al. [40], who found forests in the humid Jilin Province (600–800 mm annual precipitation) had 35% higher water conservation capacity than grasslands. However, in the semi-arid SHB region (460.3 mm annual precipitation [25]), high-density planted forests consume ~70% of precipitation through evapotranspiration, depleting soil moisture, while grasslands with lower evapotranspiration rates demonstrate superior WCS performance. This illustrates how climate regulates the WCS functions of different land-use types. In addition, grasslands typically process more favorable soil structures and better aeration than forest areas, which promote water infiltration, increase soil saturated hydraulic conductivity, and enhance soil water retention capacity [8,41]. Additionally, their well-developed root systems contribute to soil stabilization and erosion reduction, further improving WCS effectiveness [42]. These advantages can be partly attributed to regional bioclimatic conditions. As Liu et al. [1] noted, afforestation in SHB has led to rapid vegetation cover but also high water demand, decreasing water-use efficiency and limiting WCS function. Zeng et al. [12] further reported that in semi-arid grasslands, a 10% decline in vegetation cover may result in an 18 mm decrease in WCS, significantly higher than the 9 mm loss observed in this study. This discrepancy highlights the need for stricter control of grazing and tourism activities in SHB to sustain the water conservation capacity of its grassland ecosystems.

Moreover, WCS functions in cropland and built-up land were significantly lower than those of other land-use types in the SHB region (Figure 5 and Table 2), likely due to intensive human practices, such as cultivation, irrigation, and fertilization in cropland, while ground hardening in built-up areas negatively impacts WCS functions [8,41,42]. The conversion of grassland to built-up land was a key driver of WCS decline under the ED scenario compared to the EP and ND scenarios. Land-use trends from 2002 to 2022 reflected both policy interventions and environmental responses. Cropland initially expanded, then declined, and eventually stabilized under the simulated 2035-EP scenario (Figure 3), consistent with the “Master Plan for the Saihanba Mechanical Forest Farm, Hebei Province (2017–2030)”, which aims to maintain cropland stability. Forest land first decreased and then increased, with a continued rise projected under the 2035-EP scenario, reflecting successful afforestation and restoration efforts [43]. Grassland expanded sharply from 2002 to 2012 but declined afterward, a trend expected to continue through 2035, possibly due to climate change, land management practices, and conversion to forest land [44,45]. Wetland area increased slightly before returning to 2002 levels by 2022, mainly due to the conversion into forest and grassland after 2012 [46] (Figure 4 and Table 1). These fluctuations may also be attributed to local wetland protection measures and climate change [46]. Notably, under the simulated 2035-EP scenario, wetlands are projected to increase again, approaching 2002 levels, suggesting the effectiveness of land-use policies in restoring WCS functions. Concurrently, built-up land expanded steadily, mirroring ongoing urbanization, while unused land decreased, reflecting more intensive land use and regional development planning [47].

From 2002 to 2022, the WCS capacity of the SHB region initially declined and then recovered (Figure 5), closely linked to changes in land use and evolving management strategies. Specifically, to improve the ecological environment, the SHB region experienced large-scale afforestation practices from 2002 to 2012 [1,25]. However, excessive afforestation resulted in wetland degradation, weakening the region’s original WCS capacity, particularly its hydrological regulation functions [1,25]. Since 2012, the forestry management policy of the SHB region has shifted from a focus on large-scale afforestation to a more diversified and comprehensive ecological management approach [1,25], which may be the primary driver behind the recovery of WCS capacity in the region.

By 2035, regional WCS capacity varied across the three simulated scenarios (Figure 4 and Figure 5), likely due to differences in land-use patterns. Under the 2035-EP scenario (Figure 5d), the large-scale transition of forest and unused land into wetlands and grasslands led to more effective WCS optimization than under the 2035-ND and 2035-ED scenarios. However, the 2035-ND and 2035-ED scenarios (Figure 5e,f) involved substantial expansion of built-up areas at the expense of forest and unused land, with only limited restoration to wetlands and grasslands, resulting in markedly lower WCS benefits. Overall, the 2035-EP scenario exerted a stronger influence on land-use structure by expanding ecological land, curbing urban sprawl, and enhancing the spatial configuration and WCS capacity of the SHB region.

In fact, LULC not only affects regional WCS capacity but also other ESs. Substantial land-use changes can alter ecosystem structure and function, thereby shaping ES outcomes. Existing studies have shown that under the EP scenario, regional ESs such as net primary productivity, carbon storage, and soil conservation are enhanced [48,49]. In contrast, under ED scenarios, human activities, such as urbanization, result in inefficient land-use structures that diminish key ESs, including carbon storage, food supply, and water conservation, posing potential ecological risks [50,51,52]. Therefore, the EP scenario emerges as the most effective strategy for optimizing land-use structure and curbing ES degradation.

4.2. Optimization of Spatial Patterns for WCS Capacity

To further refine the zoning strategies for the simulated EP, ND, and ED scenarios by 2035, a BBN model was constructed to integrate factors influencing the status of WCS capacity. The model revealed that key variables were consistent across all three scenarios, with similar spatial patterns observed for the key state subsets (Supplementary Figure S1). Subsequently, the BBN results were divided into five subsets (Figure 7).

Subset I was identified as the region with the highest probability of poor WCS capacity (Figure 7). This subset, which is primarily distributed in the Sandaohekou Forest Farm and the western part of the Qiancengban Forest Farm, is over 95% forested. The dense canopy reduces surface runoff and water yield, while large expanses of young and near-mature forests consume substantial water during growth. Therefore, Subset I was designated as a key optimization zone (Figure 8a). In contrast to Subset I, Subset V exhibited the highest probability of strong WCS capacity, which was mainly distributed in the Sanxiang and central-eastern Yinhe Forest Farms in the SHB region. These areas within Subset V were characterized by a mix of land-use types, including forest, grassland, and wetland, and received high annual precipitation, surface runoff, and water yield. Thus, Subset V was designated as an ecological protection zone (Figure 8b), where development and industrial activities should be restricted to preserve its high WCS capacity and superior natural conditions. Areas outside the key optimization and ecological protection zones were classified as general management zones.

Compared to the investigation conducted in the Weihe River Basin, China, by Zeng et al. [12], our work not only identifies the key driving factors of WCS using BBN but also simulates multi-scenario land-use projections for 2035 by integrating the CA-Markov model. This approach achieves dynamic coupling of “scenario prediction—functional assessment—spatial optimization” and clarifies optimization directions under different development pathways through a three-step framework (Figure 2). To our knowledge, this dynamic approach represents the first application of its kind in ES research within the study area [19,52].

Furthermore, the EP scenario emerged as the optimal mode for refining the regional land-use structure and mitigating the degradation of WCS functions from an ecological benefit perspective. Therefore, the EP scenario should be implemented in conjunction with the land-use direction outlined in the “Master Plan for the Mechanical Forest Farm of Saihanba, Hebei Province (2017–2030)” to achieve a balanced land-use structure by increasing grassland and wetland while reducing other land-use types. This strategy is similar to the farmland protection scenario in the Changji-Tumen region, China [52], where restricting the conversion of farmland and limiting the expansion of built-up land effectively preserved farmland and forest land, maintaining a high local ES value. Additionally, an ecosystem restoration plan should be implemented for the areas in the SHB region affected by adverse factors, along with establishing a long-term ecological monitoring system to enable rapid responses to environmental changes.

An accurate understanding of trends in each driving factor is essential for developing effective water protection strategies [53]. Sensitivity analysis identified precipitation as the primary determinant of WCS capacity, particularly during dry seasons when fluctuations exerted significant impacts (Supplementary Table S6). This finding aligns with Wang et al. [54], who demonstrated at a global scale that precipitation is a key driver enhancing the synergy of multiple ecosystem services, including water conservation, soil retention, and climate regulation. In the semi-arid SHB region, where precipitation is relatively low, its role in sustaining WCS capacity is even more pronounced. This suggests that even minor changes in precipitation could have substantial effects on WCS capacity. Additionally, land-use changes, especially the conversion of wetlands and/or forests to other types, significantly affected WCS capacity. These findings underscore the necessity of integrating precipitation, land-use changes, and other driving factors into water resource protection and land-use planning to enhance WCS capacity effectively.

In summary, building eco-friendly environments to support the WCS capacity of the SHB region is crucial. Strict protection approaches should be implemented in ecologically significant and fragile areas, particularly in key biodiversity habitats, to prevent environmentally destructive development. Water source protection and water quality management should be prioritized in land-use planning to ensure water resource security in forest farms and surrounding areas. Furthermore, it is recommended to establish nature reserves in the Sandaohekou, Beimandian, Yinhe, and Qiancengban Forest Farms in the SHB region, as well as to create ecological protection forests in the transitional mountainous areas of the Qiancengban and the Dahuanqi Forest Farms. Commercial timber production bases should be established in the five Mandian areas (lava landforms) of the Dahuanqi and Sanxiang Forest Farms. Additionally, the development of oasis ecological construction should be actively promoted. Establishing artificial grasslands will restore the ecological environment and substantially enhance the water-conservation capacity of optimization zones. Zuo et al. [55] reported that in the northern Qilian Mountains, the gradual conversion of pure coniferous plantations into mixed conifer–broadleaf forests can increase WCS capacity per unit area by approximately 19%. This finding suggests that the key optimization zones identified in this study should prioritize stand structure regulation measures. Land-use planning should be conducted in a scientifically sound and regulatory-compliant manner, considering the specific conditions of the forest farm, protecting land resources, and balancing ecological protection with regional economic development. Consequently, the WCS optimization zone in the SHB region should prioritize the Sandaohekou Forest Farm and the western part of the Qiancengban Forest Farm for enhanced greening projects (Figure 8).

4.3. Limitations and Future Directions

The BBN is subject to limitations stemming from structural, input data, and parametric uncertainties [56]. To mitigate structural uncertainty, future research could incorporate additional factors, such as socioeconomic variables, to enhance the understanding of WCS processes [56]. Peng et al. [39] highlighted the significance of socio-economic factors, including tourism pressure, in ecological restoration decisions, as these influence land-use changes and ecosystem service optimization. Therefore, future research should refine the BBN model by incorporating more socio-economic variables to improve its real-world applicability and support policymaking. Li et al. [43] proposed an evaluation framework for nature-based solutions, emphasizing the need to balance ecological conservation with economic development. While tourism can negatively impact ecosystem services, it also generates economic benefits. Therefore, policies should strive for a balanced approach that integrates effective planning to ensure both environmental protection and sustainable economic growth.

Additionally, the CA-Markov model has certain limitations in long-term predictions. Chen et al. [57] noted that LULC models often struggle to account for stochastic disturbances, such as extreme climate events or sudden policy shifts, which can significantly impact land-use changes. Due to its reliance on historical data and deterministic rules, the CA-Markov model is less effective in simulating such uncertainties. Therefore, when applying this model for long-term projections, caution is required, and complementary approaches, such as scenario analysis, should be considered to better address uncertainties.

Moreover, new technologies, such as unmanned aerial vehicle RS, could be employed to obtain high-precision spatial data, thereby enhancing the model’s ability to represent spatial heterogeneity and enabling a more comprehensive evaluation of the synergies and trade-offs between WCS and other ESs. Such advancements would support the development of comprehensive ecological management strategies [58]. Additionally, strengthening collaboration with local governments and communities is crucial to promote the implementation of optimization approaches and achieve sustainable utilization of ES [49]. Finally, encouraging interdisciplinary integration will facilitate a deeper analysis of optimization challenges in the SHB region, contribute to the establishment of long-term monitoring systems, and aid in assessing the impacts of climate change on WCS, thereby supporting regional sustainable development.

5. Conclusions

As a key ecological zone historically degraded by excessive logging, the SHB region requires targeted strategies to enhance its water conservation functions. In this study, the WCS capacity in the SHB region was quantitatively evaluated under multiple LULC scenarios using the CA-Markov and BBN models. BBN models effectively handle multiple sources of uncertainty using CPTs and Bayesian inference, making them particularly advantageous for land-use change and ES optimization, especially in the context of complex multifactor interactions. Key findings include the following: (1) LULC significantly impacted WCS capacity from 2002 to 2035, with forests and grasslands as dominant land-use types and urban lands as the least effective. The EP scenario yielded optimal WCS status by 2035. (2) WCS capacity declined and then recovered, with higher values in the southern and central-eastern parts of the SHB region, especially under the EP scenario. (3) BBN optimization identified three distinct zones: key optimization (15% of the SHB region), ecological protection (30%), and general management (55%). The key optimization zone, mainly in Sandaohekou Forest Farm and the western part of the Qiancengban Forest Farm, requires attention due to high forest density and soil degradation risks. The ecological protection zone in Sanxiang and central-eastern Yinhe Forest Farms benefits from a balanced distribution of wetlands, grasslands, and forests. The general management zone serves as a transitional area, maintaining moderate WCS capacity. These findings offer practical guidance for improving land-use planning, enhancing ecological resilience, and ensuring long-term water security. Future efforts should aim to improve monitoring efficiency and promote ecosystem sustainability through the integration of RS, GIS, and machine learning techniques.