Integrating Knowledge Graphs and Bayesian Inference to Balance Ecological Security, Carbon Sinks, and Development: A Case Study of Land Use Zoning in Yunnan

Abstract

Balancing ecological protection, carbon sinks, and development is a practical challenge in mountainous regions. Using Yunnan Province, China, as a case study, this paper develops a knowledge-guided probabilistic framework for carbon-oriented territorial zoning. The framework combines an indicator system, corridor analysis of pattern, risk and potential, knowledge-graph rule encoding, Bayesian mechanism calibration, and constrained posterior decoding on 11,853 effective planning cells. The results show a clear conservation–development gradient in the carbon sink priority surface: high-priority areas are concentrated in western and southwestern Yunnan, whereas low-priority areas cluster around major urban centers. Corridor analysis identifies a central resistance belt and several urban–rural bottlenecks, indicating that connectivity constraints are concentrated in a limited number of critical links. The final zoning assigns 35.4% of grids to integrated development, 25.9% to emergency intervention, 14.5% to long-term conservation, 13.8% to priority restoration, and 10.4% to risk control. Zone separability is generally strong, with one-versus-rest AUC values ranging from 0.777 to 0.995. Land use enrichment further supports the zoning results: integrated development contains 78.85% of built-up land and 45.93% of cropland, whereas Emergency intervention, priority restoration, and long-term conservation together contain 70.01% of forest area.

1. Introduction

Reconciling ecological security, climate change mitigation, and socioeconomic development has emerged as a central challenge for territorial governance worldwide [1,2]. In the context of China’s carbon neutrality goals, the need to spatially coordinate carbon sink conservation and enhancement with other policy objectives has become particularly acute, especially in mountainous provinces that serve as critical ecological barriers and carbon pools [3,4]. Yunnan Province, located in Southwest China, exemplifies this challenge: it harbors some of the country’s most extensive forest ecosystems and biodiversity reserves, yet simultaneously faces accelerating urbanization, infrastructure expansion, and land use intensification that fragment habitats and elevate ecological risk. Designing spatially explicit zoning schemes that can guide differentiated management interventions, ranging from emergency restoration to compatible development, is therefore essential for operationalizing carbon-oriented territorial governance in such complex social–ecological systems [5].

Substantial methodological progress has been made in spatial conservation prioritization and ecological zoning. Core Area Zonation based on marginal loss ranking has proven effective for identifying high-priority areas for biodiversity and ecosystem services by iteratively removing cells that contribute least to conservation value [6,7]. Complementing this, circuit theory applied to resistance surfaces enables characterization of ecological corridors and bottleneck locations, with corridor efficiency quantified through metrics such as cost distance and the cost-weighted distance-to-path-length ratio [8]. These approaches provide valuable biophysical evidence for where ecological assets are concentrated and where functional connectivity is most vulnerable. However, translating such diagnostics into implementable zoning decisions requires navigating a complex policy space defined by multiple objectives, governance thresholds, and scenario uncertainties, and a task for which existing methods remain inadequately equipped [9]. Compared with ecosystem service-based planning frameworks that map and value ecosystem functions to guide land use decisions [10], the proposed approach complements these methods by explicitly incorporating governance rules and uncertainty through knowledge graphs and Bayesian modeling. This enables not only the identification of priority areas, but also the generation of implementable and auditable zoning decisions under competing objectives.

Methodologically, many zoning studies continue to rely on deterministic multi-criteria scoring and heuristic optimization, including multi-criteria decision analysis and evolutionary algorithms [11,12,13]. These approaches are useful for exploring trade-offs, but often lack rigorous uncertainty quantification and are difficult to audit against explicit governance thresholds [14,15]. The resulting maps may satisfy mathematical optimality criteria but fail to align with regulatory feasibility or to communicate the confidence associated with each allocation. Moreover, the mechanisms linking landscape structure, ecological risk, and carbon dynamics are typically treated implicitly, limiting the capacity for process-grounded inference and scenario-based projection [16,17,18]. Despite substantial progress in conservation prioritization, corridor analysis, and land use optimization, territorial zoning studies still rarely integrate three elements within a single auditable workflow, which are spatial ecological evidence, explicit governance rules, and formal uncertainty quantification. Existing approaches usually perform well in one or two of these dimensions, which are ecological diagnosis, heuristic allocation, or policy interpretation. However, they seldom integrate these dimensions in a manner that is simultaneously process-informed, rule-feasible, and probabilistically transparent. This gap is especially consequential in mountainous provinces, where ecological resources, development pressure, and governance constraints are sharply uneven in space and where planning decisions must reconcile restoration urgency, long-term conservation, and development demand under uncertainty [19,20].

To address these gaps, this study developed a four-stage, knowledge-guided probabilistic zoning framework for carbon-oriented territorial planning. First, PRP-aligned ecological evidence is constructed on a unified grid using indicators of connectivity, ecological risk, carbon sink, and development pressure. Second, governance knowledge is encoded in an ecological governance knowledge graph, while a Bayesian structural equation model provides mechanism-informed prior calibration. Third, a Bayesian mixed-effects multinomial model estimates grid-level posterior probabilities for five management zones while accounting for spatial heterogeneity Fourth, the final zoning map is obtained through constrained posterior decoding, ensuring that each allocation maximizes probabilistic support while satisfying all knowledge graph feasibility rules.

The objective of this study is to establish a rule-consistent, uncertainty-aware spatial decision framework that integrates knowledge graphs with Bayesian inference to address the coordination between carbon sink conservation, ecological security, and development pressures in mountainous territorial governance, using Yunnan Province as a case study. The framework contributes in three principal ways. First, it operationalizes a pattern–risk–potential (PRP) indicator system to establish an evidence base for territorial differentiation. Second, it combines knowledge graph-encoded governance semantics with Bayesian structural equation modeling that explicitly links ecological resistance, connectivity, risk, and carbon potential, thereby strengthening the process-grounded basis for zoning decisions. Third, it delivers an auditable and updateable zoning scheme through knowledge graph-constrained Bayesian mixed-effects learning and constrained posterior decoding. The framework is designed to support monitoring, adaptive management, and rolling revisions within evolving territorial governance systems.

2. Study Area and Data Sources

2.1. Study Area

Yunnan Province (21°08′32″–29°15′08″ N, 97°31′39″–106°11′47″ E) is located in Southwest China, covering a total area of 394,000 km². It spans strong elevational and climatic gradients from the high-relief northwest to the lower-elevation southeast, making it a representative mountainous region where development corridors and ecological security functions are spatially intertwined. Yunnan lies at the convergence of the Mountains of Southwest China, the Eastern Himalayas, and Indo-Burma, and therefore plays a critical role in safeguarding ecological connectivity and ecosystem services at regional and transboundary scales (Figure 1).

2.2. Data Sources and Preprocessing

We integrated multi-source geospatial and statistical datasets to characterize the pattern–risk–potential indicators and support subsequent zoning analysis, as summarized in

Table 1. Summary of data sources and source information.

Data Type	Data Name	Data Source	Explanation
Vegetation	Forest Net Primary Productivity (MOD17A3H V6)	NASA Earth Observing System Data and Information System (https://search.earthdata.nasa.gov/ (accessed on 16 November 2025))	500 m resolution, NPP product for carbon sink assessment of 2025
Land Use Data Physical Geographic Data	Land Use	Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences (https://www.resdc.cn (accessed on 16 November 2025))	Land use types for 2025, 30 m resolution
	Digital Elevation Model (DEM)	Geospatial Data Cloud, Chinese Academy of Sciences (http://www.gscloud.cn/ (accessed on 16 November 2025))	SRTMGL1_003, 30 m resolution
	Slope	Derived from DEM	30 m resolution, derived from elevation data
Socioeconomic Data	Normalized Difference Vegetation Index (NDVI)	Geospatial Data Cloud (https://www.gscloud.cn (accessed on 30 November 2025))	Derived from red and near-infrared band spectra
	Gross Domestic Product (GDP) per unit area	Resource and Environmental Science Data Center, Chinese Academy of Sciences (https://www.resdc.cn (accessed on 30 November 2025))	1 km resolution
	Population Density (POP)	WorldPop (https://www.worldpop.org/ (accessed on 30 November 2025))	Number of people per unit area of land, 1 km resolution ERA5 reanalysis dataset, 0.1° resolution
Climate Data	Annual Average Precipitation	ECMWF (https://www.ecmwf.int/ (accessed on 30 November 2025))
	Annual Average Temperature	ECMWF (https://www.ecmwf.int/ (accessed on 30 November 2025))
Administrative Boundaries	Urban Administrative Boundary Data	National Platform for Common GeoSpatial Information Services (https://cloudcenter.tianditu.gov.cn/dataSource (accessed on 30 November 2025))	Administrative division map of Yunnan Province

.

All raster datasets were resampled or aggregated to 1 km × 1 km grid resolution to ensure spatial alignment. Considering the need to preserve regional heterogeneity while maintaining computational tractability and stable inferential performance, we retained 6 km as the primary planning unit (Table S1). This resolution balances the need to capture landscape heterogeneity with the computational demands of the Bayesian mixed-effects multinomial model applied across the entire province. The effective planning grid comprising 11,853 cells serves as the fundamental decision unit for subsequent zoning analysis. All indicators, which are Ecological Security Pattern Connectivity, Ecological Risk Index, Carbon Sink Priority, and GDP intensity, were standardized to zero mean and unit variance to ensure cross-dimensional comparability and to facilitate probabilistic learning within the Bayesian modeling framework. For knowledge graph rule encoding, indicators were min–max-normalized to the [0, 1] interval so that threshold semantics retained direct interpretability. For Bayesian estimation, the same indicators were z-standardized to zero mean and unit variance, so model coefficients represent the effect of a one-standard-deviation change in a predictor on the log-odds of zone assignment.

3. Materials and Methods

3.1. Research Framework

The workflow proceeds in four steps (Figure 2). First, pattern, risk and potential (PRP) indicators are quantified on a common scale. Each effective grid cell is characterized by four standardized indicators forming a PRP-aligned evidence system, with GDP intensity added to represent development pressure. Ecological Security Pattern Connectivity (ESPC) represents spatial ecological structure, Ecological Risk Index (ESI) captures disturbance and vulnerability, carbon sink priority (CSP) serves as a productivity-driven sequestration potential index reflecting intervention-oriented carbon gain leverage, and GDP intensity measures development pressure. Second, rule and prior encoding is performed via ecological governance knowledge (KG). Governance knowledge is formalized as auditable semantics for five management zones, which are emergency intervention (EI), risk control (RC), priority restoration (PR), long-term conservation (LC), and integrated development (ID). The auditable semantics for five management zones are translated into two complementary components: one is hard feasibility constraints that define the KG-feasible candidate zone set for each grid cell, and the other is soft prior guidance that determines directional and magnitude priors for model parameters. Third, zoning is learned via a Bayesian mixed-effects multinomial model. This approach estimates grid-level posterior probabilities for the five zones. Fourth, the final zoning allocation is obtained via Constrained Posterior Decoding (CPD), which selects the most supported zone within each grid cell’s KG-feasible domain. The resulting zoning outcomes are then validated through indicator contrasts, land use composition, concentration diagnostics, and supplemented by stratified visual validation using very-high-resolution imagery.

3.2. PRP Indicator Construction

The PRP indicator set was designed to preserve dimensional parsimony while covering the four core processes required for carbon-oriented territorial zoning: ecological structure (ESPC), disturbance and vulnerability (ESI), sequestration leverage (CSP), and development pressure (GDP intensity). Additional variables commonly used in ecological zoning, such as road density, fragmentation, or edge pressure, were not ignored; rather, they were incorporated upstream in the construction of composite dimensions, especially risk and connectivity, to avoid semantic redundancy and excessive collinearity among top-level zoning predictors. Pattern is quantified by Ecological Security Pattern Connectivity (ESPC) derived from ecological sources and functional corridors. Ecological sources are identified using morphological/spatial pattern methods and MSPA-based core extraction [21]. Corridors and pinch points are characterized using circuit theory on an integrated resistance surface. Corridor efficiency is summarized using complementary metrics, including the least-cost distance (lcDist) and the cost-weighted distance-to-path-length ratio:

$$\mathrm{c}\mathrm{w}\mathrm{d}\mathrm{T}\mathrm{o}\mathrm{P}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{R}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}={\displaystyle \frac{\mathrm{C}\mathrm{W}\mathrm{D}}{\mathrm{L}\mathrm{C}\mathrm{P}\_\mathrm{L}\mathrm{e}\mathrm{n}\mathrm{g}\mathrm{t}\mathrm{h}}}$$

(1)

where larger values indicate stronger barriers even when geometric distances are limited. Effective resistance (Eff_Resist) is used as an integrated measure of functional pathway impedance [22].

Risk is quantified using an Ecological Risk Index (ESI) computed on a 6 km × 6 km fishnet to integrate landscape structure and land use vulnerability. This resolution balances the need to capture landscape heterogeneity with the computational demands of the Bayesian mixed-effects multinomial model applied across the entire province. Landscape metrics are computed by Fragstats, combined with land use vulnerability scores, and then interpolated to the planning grid using geostatistical methods to obtain a continuous risk surface [23].

Potential is represented by carbon sink priority (CSP). Vegetation net primary productivity (NPP) is a critical indicator for assessing ecosystem carbon sink capacity and sustainability, and therefore serves as the foundational metric for prioritizing areas where carbon gains can most effectively be achieved [24,25]. In this study, CSP is operationalized as a zonation-derived priority score ranging from 0 to 1, driven exclusively by multi-year mean NPP as a productivity-driven proxy for sequestration potential. The resulting priority score is interpreted as intervention leverage for potential carbon gains, and is standardized before entering the probabilistic zoning model [26].

Development pressure is represented by GDP intensity, defined as economic output per unit area [27].

3.3. Bayesian Structural Equation Model

The Bayesian SEM is used as a mechanism–calibration layer to encode directionally informed priors for the multinomial zoning model based on empirically supported pathways among resistance, connectivity, risk, and carbon potential. To characterize causal pathways among spatial structure, ecological risk, productivity-driven sequestration potential, and economic pressure, and to maintain consistency with county-level empirical evidence, we construct a Bayesian structural equation model (Bayesian-SEM) as a mechanism interpretation layer [28,29]. Gaussian errors are adopted to accommodate heteroscedasticity and outliers; Student’s t likelihood can be used as a robust extension without changing model structure [30].

A “resistance–connectivity–carbon” pathway is specified, reflecting empirical evidence that increased resistance reduces connectivity and carbon advantage, while economic pressure elevates risk and suppresses connectivity:

$$\begin{array}{ll}& {\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{n}}_{\mathrm{c}}\sim \mathrm{N}({\delta }_0+{\delta }_1{M}_c+{\delta }_2{E}_c,{\sigma }_C^2),\\ & {\mathrm{R}}_{\mathrm{c}}\sim \mathrm{N}({\rho }_0+{\rho }_1{E}_c-{\rho }_2{C}_c,{\sigma }_R^2),\\ & S_{\mathrm{c}}\sim \mathrm{N}({\gamma }_0+{\gamma }_1{C}_c-{\gamma }_2{R}_c-{\gamma }_3{E}_c,{\sigma }_S^2).\end{array}$$

(2)

where Mc denotes resistance, Cc represents connectivity, Ec denotes economic pressure, Rc is ecological risk, and Sc is carbon advantage. Expected signs are δ1 < 0, γ1 > 0, and γ2 > 0. To align the mechanism with this evidence, weakly informative or informative priors are specified:

$${\mathsf{\delta }}_1\sim \mathrm{N}({\mathsf{\mu }}_{\mathsf{\delta }1},{\mathsf{\sigma }}_{\mathsf{\delta }1}^2),{ \mathsf{\mu }}_{\mathsf{\delta }1}<0, {\gamma }_1\sim {\mathrm{N}}^{+}(0,1)$$

(3)

where N+ denotes a truncated normal distribution. The magnitude of |μ_δ1| is calibrated using the absolute value of the empirical slope, with a negative sign reflecting “higher resistance to lower carbon advantage”. Other coefficients follow zero-mean weakly informative normal priors [31].

3.4. Rule Encoding and Prior Injection

3.4.1. Knowledge Representation

To inject auditable governance knowledge into data-driven learning, an ecological governance knowledge graph (KG) is constructed and used to define hard constraints and soft prior guidance. Knowledge is encoded as triples [32]:

$$\begin{array}{c}\left(\mathrm{h},\mathrm{r},\mathrm{t}\right)\in \mathcal{E}\times \mathcal{R}\times \mathcal{E},\mathrm{h},\mathrm{t}\in \mathcal{E},\mathrm{r}\in \mathcal{R}.\end{array}$$

(4)

where entities ε include grid units, corridors and pinch points, indicators (ESPC, ESI, CSP, GDP), zoning types, and governance actions; relations R include promote/inhibit, suitable-for, constrain, located-in, and connect. Zoning types are defined as emergency intervention (EI), risk control (RC), priority restoration (PR), long-term conservation (LC), and integrated development (ID). The knowledge graph was assembled from three sources: indicator semantics defined by the PRP framework, zoning objectives from territorial–governance practice, and threshold-based feasibility rules used to determine admissible zone assignments. Each rule was encoded as a subject–relation–object triple and linked to one of two model functions: sign and magnitude control in prior specification, or cell-wise feasibility restriction in constrained posterior decoding (Table S2).

3.4.2. Directional and Magnitude Priors

KG rules determine coefficient signs and prior means. For compactness, priors are expressed as follows:

$$\begin{array}{c}{\mathsf{\beta }}_{\mathrm{k},\mathrm{m}}\sim \mathrm{N}\left(\mathrm{s}\mathrm{g}{\mathrm{n}}_{\mathrm{k},\mathrm{m}}\cdot {\mathsf{\lambda }}_0{\mathrm{w}}_{\mathrm{k},\mathrm{m}},{\mathsf{\sigma }}_{\mathsf{\beta }}^2\right)\end{array}$$

(5)

where k∈{1, …, 5} indexes zoning types and m indexes indicators, w_k,m are entropy-based indicator weights, sgn_k,m∈{−1, 0, +1} is determined by KG rules, and λ₀ is a scaling factor. Accordingly, relations such as promote, inhibit, and suitable-for were translated into sgn_k,m, prior magnitudes weighted by w_k,m, and admissible candidate sets used in decoding.

3.5. Knowledge Graph-Constrained Bayesian Mixed-Effects Multinomial Model

To replace deterministic scoring and Pareto-based optimization at the zoning stage, zoning is modeled as a probabilistic multi-class decision under spatial heterogeneity using a Bayesian mixed-effects multinomial model [33]. For grid g_i, the zoning label z_i∈{1, …, 5} follows a categorical distribution:

$$\begin{array}{c}{\mathrm{z}}_{\mathrm{i}}\sim \mathrm{C}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{g}\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{a}\mathrm{l}\left({\mathrm{p}}_{\mathrm{i}}\right),{\mathrm{p}}_{\mathrm{i}\mathrm{k}}=\mathrm{P}\mathrm{r}({\mathrm{z}}_{\mathrm{i}}=\mathrm{k})\end{array}$$

(6)

A softmax link with mixed effects is used:

$$\begin{array}{c}{\mathrm{p}}_{\mathrm{i}\mathrm{k}}={\displaystyle \frac{\exp ({\mathsf{\eta }}_{\mathrm{i}\mathrm{k}})}{\sum _{\mathrm{j}=1}^5\exp ({\mathsf{\eta }}_{\mathrm{i}\mathrm{j}})}},{\mathsf{\eta }}_{\mathrm{i}\mathrm{k}}={\mathsf{\alpha }}_{\mathrm{k}}+{\mathsf{\beta }}_{\mathrm{k}}^{\mathrm{T}}{\tilde{\mathrm{x}}}_{\mathrm{i}}+{\mathrm{u}}_{\mathrm{c}\left(\mathrm{i}\right),\mathrm{k}}+{\mathrm{v}}_{\mathrm{s}\left(\mathrm{i}\right),\mathrm{k}}\end{array}$$

(7)

where x_i = (ESPCi, ESIi, CSPi, GDPi)⊺. The random effects u_c,k and v_s,k capture county-level spatial heterogeneity.

3.6. Constrained Posterior Decoding for Final Zoning

To convert probabilistic predictions into implementable zoning maps, we apply constrained posterior decoding (CPD), which selects zoning assignments maximizing posterior probability subject to governance constraints [34,35].

3.6.1. Objective Function

The maximum a posteriori allocation is as follows:

$$\begin{array}{c}\underset{\left\{{\mathrm{z}}_{\mathrm{i}}\right\}}{\max }\sum _{\mathrm{i}=1}^{\mathrm{N}}\log {\mathrm{p}}_{\mathrm{i},{\mathrm{z}}_{\mathrm{i}}}\end{array}$$

(8)

Optionally, a linear utility term can encode multi-objective preferences:

$$\begin{array}{c}\underset{\left\{{\mathrm{z}}_{\mathrm{i}}\right\}}{\max }\sum _{\mathrm{i}=1}^{\mathrm{N}}\left(\log {\mathrm{p}}_{\mathrm{i},{\mathrm{z}}_{\mathrm{i}}}+\mathsf{\lambda }{\mathrm{U}}_{\mathrm{i},{\mathrm{z}}_{\mathrm{i}}}\right),{\mathrm{U}}_{\mathrm{i}\mathrm{k}}={\mathrm{w}}_{\mathrm{k}}^{\mathrm{T}}{\tilde{\mathrm{x}}}_{\mathrm{i}}\end{array}$$

(9)

where λ controls the trade-off between posterior support and user-specified utility, w_k is the weight vector for zone k, and x_i is the PRP indicator vector.

3.6.2. Constraints

Hard feasibility constraints enforce KG rules by restricting each grid cell to its admissible zone set:

$$\begin{array}{c}{\mathrm{z}}_{\mathrm{i}}\in {\mathcal{K}}_{\mathrm{i}},\forall \mathrm{i}\end{array}$$

(10)

where K_i is the cell-wise admissible zone set derived from KG threshold rules. Optional area constraints ensure planning-scale consistency:

$${\mathrm{L}}_{\mathrm{k}}\le {\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{a}}_{\mathrm{i}}\mathbb{I}\left({\mathrm{z}}_{\mathrm{i}}=\mathrm{k}\right)\le {\mathrm{U}}_{\mathrm{k}},\mathrm{k}=1,\dots ,5,$$

(11)

where N is the number of effective grid cells, ai is the effective area of cell i, $\mathbb{I}({\mathrm{z}}_{\mathrm{i}}=\mathrm{k})$ is an indicator function that equals 1 if cell i is assigned to zone k and 0 otherwise, and L_k and U_k are user-specified lower and upper bounds on the total area allocated to zone k.

The Bayesian SEM, knowledge-graph scoring, surrogate-based probabilistic zoning, and fusion decoding were implemented in Python (v3.10) using NumPy, pandas, scikit-learn, and SciPy.

4. Results

4.1. Carbon Sink Prioritization Delineates a Conservation–Development Gradient

To evaluate how conservation value is retained along the priority ranking, we examined performance curves derived from the marginal loss rule together with the feature histogram (Figure 3). As priority scores increase from 0 to 1, the curve for positively weighted (carbon-beneficial) features shows a gradual decline, indicating that these features are consistently captured within high-priority areas and are lost only progressively as the solution expands into lower-priority space. The relatively gentle slope further suggests that carbon-beneficial features are not confined to a few isolated hotspots but are spatially widespread, and that the prioritization effectively secures their core distributions early in the ranking. In contrast, the curve for negatively weighted features decreases much more steeply, implying that these features are disproportionately concentrated toward the low-priority tail. Consistently, the histogram of the top priority fraction indicates that the upper portion of the ranking is dominated by beneficial features, whereas undesirable features are under-represented and accumulate primarily outside the high-priority envelope.

The resulting carbon sink priority surface exhibits a coherent and interpretable spatial gradient from ecologically intact vegetated landscapes toward ecologically vulnerable areas and urban cores. The highest-priority fraction (top 80–100%) is concentrated in southwestern and western Yunnan and around major plateau lakes, whereas the lowest-priority areas cluster around major cities and overlap with ecologically constrained landscapes. Land use composition corroborates this gradient: forest dominates the highest-priority class and declines monotonically with decreasing priority, while cropland and grassland increase progressively in lower-priority categories. Built-up land is most concentrated in the lowest-priority class, reinforcing the functional interpretation that low-priority areas represent candidate spaces for development and/or more intensive management, whereas high-priority areas represent the core geography where carbon sink protection is both ecologically grounded and most efficiently targeted by the weighting scheme.

4.2. Corridor Efficiency Diagnostics Identify a Central Resistance Belt and Urban–Rural Bottlenecks

Circuit theory connectivity mapping reveals pronounced spatial heterogeneity in corridor structure and movement impedance across Yunnan (Figure 4). A clear “central resistance belt” emerges in central Yunnan, where functional connections between source patches are stretched over longer inter-patch distances and further constrained by built-up expansion, exposed/bare surfaces, and rugged terrain. In these areas, corridors are not only spatially elongated but also traversed through landscapes with persistently elevated resistance, indicating that maintaining connectivity relies on a limited set of costly pathways. Bottleneck locations cluster around urban–rural transition zones, where rapid land use turnover and fragmentation compress ecological flows into narrow passages, making the network locally vulnerable to disruption. By contrast, the southwestern region exhibits comparatively lower effective resistance and more continuous corridor linkages, indicating more favorable conditions for maintaining a coherent ecological network. The southeastern karst region shows sparse corridor connections, consistent with its fragmented habitat configuration and a limited set of structurally viable links.

These patterns indicate that connectivity in Yunnan is broadly maintainable but highly uneven in space. Across all corridor links, lcDist, cwdToPathRatio, and Eff_Resist remain moderate for most segments. In particular, high-cost and low-efficiency segments are concentrated mainly in central Yunnan and around urban–rural interfaces, where ecological flows are compressed into narrow passages and are therefore more vulnerable to land use disturbance. This suggests that the most urgent planning priority is not uniform corridor expansion, but the targeted protection and restoration of a limited number of structurally critical links.

4.3. Knowledge Graph-Constrained Probabilistic Zoning Produces Rule-Feasible and Interpretable Spatial Differentiation

4.3.1. Governance Semantics Supported by Indicator Contrasts and Multi-Evidence Consistency

Zoning was implemented on a unified effective planning grid (n = 11853).

Table 2. Zone-wise mean standardized indicator profiles and one-versus-rest AUC values.

Zone	ESPC	ESI	CSP	GDP	One-Versus-Rest AUC
EI	−0.0463	−0.1451	−0.4072	0.6080	0.8585
ID	0.0784	0.2221	−0.3250	−0.0319	0.7771
LC	−0.4207	−0.9717	0.9105	−0.1894	0.9950
PR	0.0446	0.5269	0.6703	−0.1544	0.9320
RC	0.0725	−0.1971	0.4767	−0.0849	0.9156

reports zone-wise mean standardized scores for the four indicators; so, the governance semantics of each zone should be interpreted from their joint profiles rather than from any single-indicator extreme.

The updated profiles reveal clear and policy-relevant differentiation. Long-term conservation (LC) is the most distinctive ecological regime, combining the highest carbon potential (CSP is 0.9105), the lowest ecological risk (ESI is −0.9717), and below-average GDP intensity (−0.1894), which identifies it as a stable ecological core. Priority restoration (PR) combines the highest ecological risk (ESI is 0.5269) with high carbon potential (CSP = 0.6703), indicating strong restoration leverage. Risk control (RC) shows slightly above-average connectivity (ESPC is 0.0725), below-average ecological risk (ESI is −0.1971), and positive carbon potential (CSP is 0.4767), consistent with maintenance-oriented control space. Emergency intervention (EI) is characterized by the highest development pressure (GDP is 0.6080) and the lowest carbon potential (CSP is −0.4072), suggesting ecologically weak areas under strong human pressure. Integrated development (ID) exhibits slightly positive connectivity and ecological risk (ESPC is 0.0784; ESI is 0.2221), below-average carbon potential (CSP is −0.3250), and near-mean GDP intensity (−0.0319), representing mixed development landscapes under ecological constraints.

These distinctions are further supported by one-versus-rest AUC values, which measure how well each zone can be separated from the others using model-derived probability scores. AUC ranges from 0.7771 for ID to 0.9950 for LC, indicating very strong separability for LC and generally strong separability for PR (0.9320) and RC (0.9156), but greater overlap for EI (0.8585) and especially ID. Overall, conservation- and restoration-oriented zones are more sharply defined than human-dominated governance zones.

4.3.2. Knowledge Graph Feasibility Rules Enable Implementable Posterior Decoding

To enforce implementability, hard threshold constraints were encoded as knowledge graph (KG) feasibility rules that define the admissible zone set for each grid cell (

Table 3. Objectives and constraints for carbon reduction and carbon sequestration.

Zone	Zone Characterization	Optimization Focus (Objective Emphasis)	Hard Constraints
EI	Critical areas with very high risk and very high carbon value	$f_E(x)=0.18x_1+0.5x_2+0.25x_3+0.07x_4$	ESI ≥ 0.8; ESPC ≥ 0.8
RC	High-risk areas with moderate carbon potential	$f_R(x)=0.2x_1+0.4x_2+0.3x_3+0.1x_4$	0.3 ≤ ESI ≤ 0.6; 0.7 ≤ CSP ≤ 0.8
PR	High carbon potential with non-negligible risk; strong restoration leverage	$f_P\left(x\right)=0.25x_1+0.35x_2+0.2x_3+0.1x_4$	0.5 ≤ ESI ≤ 0.8; CSP ≥ 0.7 ESPC ≥ 0.8
LC	Ecologically stable areas with high sequestration potential and low risk	$f_L\left(x\right)=0.35x_1+0.2x_2+0.35x_3+0.05x_4$	ESI ≤ 0.3; CSP ≥ 0.6; 0.5 ≤ ESPC ≤ 0.8
ID	Low connectivity and limited carbon potential with strong development demand	$f_C(x)=0.1x_1+0.3x_2+0.15x_3+0.45x_4$	0.3 ≤ ESI ≤ 0.8; CSP ≤ 0.5; ESPC ≤ 0.3; GDP ≥ 0.6

). This explicit rule base removes logically infeasible allocations before decision-making and makes the governance logic auditable.

To address this limitation, we reformulated the KG constraints into a soft scoring system that evaluates the relative suitability of each zoning category for each grid cell. This transformation preserves governance semantics while allowing for flexible adaptation to empirical data distributions, thereby avoiding excessive infeasibility and improving compatibility with probabilistic modeling.

Within this rule-informed framework, Bayesian priors and mechanism constraints were translated into zone-wise probability estimation through cross-validated surrogate classification, and the final zoning was obtained by constrained posterior decoding within the KG-feasible candidate set.

The implementable allocation across effective grids is dominated by integrated development (ID has 4196 cells; 35.4%) and emergency intervention (EI has 3066 cells; 25.9%), followed by long-term conservation (LC has 1723; 14.5%), priority restoration (PR has 1640; 13.8%), and risk control (RC has 1228; 10.4%). Area-weighted shares show a similar but non-identical structure: EI accounts for 28.22% of effective area, whereas ID accounts for 30.16%. The grid area divergence is most evident for ID, whose mean effective unit area is substantially smaller (27.40 km²) than the other zones, indicating that development-carrying space tends to be distributed in smaller, more fragmented units and is more strongly truncated by masks and boundary effects. These spatial patterns are further contextualized by standardized zone profiles, which reveal underlying semantic and governance structures.

Building on these allocation patterns, the standardized zone profiles (Figure 5) demonstrate that conservation and restoration zones exhibit compact semantics, while development and intervention regimes span broader and more heterogeneous governance spaces.

4.3.3. Validation of Zoning Rationality Using Bayesian SEM and Confusion Matrix

The confusion matrix (Figure 6) substantiates the process-based interpretation through asymmetric separability among zones. Long-term conservation (LC) achieves the highest classification accuracy (>95%), consistent with its distinct mechanism signature—high carbon sink potential and low ecological risk. The Bayesian SEM path coefficients (Table S3, Panel A) confirm that GDP exerts a significant negative total effect on CSP (−0.201), while ESPC positively contributes (0.129), aligning with the ecological-economic causality assumed in the mechanism design. In contrast, emergency intervention (EI) and integrated development (ID) exhibit the highest mutual misclassification rates (approximately 28% and 22%, respectively), reflecting their overlapping mechanism profiles: both are characterized by low carbon potential and moderate-to-high economic pressure.

Mechanism-derived scores align with zoning outcomes. As detailed in Table S3 (Panel B and Panel C), LC shows the highest positive alignment, whereas EI exhibits strongly negative alignment. Together, the Bayesian SEM and confusion matrix confirm that the zoning scheme reflects underlying ecological–economic processes rather than arbitrary classification, supporting the integration of ecological and economic objectives within a coherent territorial governance framework.

4.4. Cross-Module Validation Confirms Semantic Coherence and Actionable Planning

Through enrichment/concentration patterns, we further verify that the five zones delineate distinct management targets rather than reflecting a single-indicator partition (Figure 7). Built-up land is highly concentrated in the integrated development (ID) zone. Although ID accounts for only 30.16% of the effective area, it contains 78.85% of total built-up land within the effective mask (3.69 × 10³ km² out of 4.68 × 10³ km²). Cropland exhibits a similar, though weaker concentration, with 45.93% of total cropland located in ID (31.05 × 10³ km² out of 67.60 × 10³ km²). In contrast, forests, which are the dominant carbon sink carrier, are preferentially secured by the protection restoration conservation spaces. The spaces are the combined EI, PR, and LC zones, which encompass 70.01% of total forest area (153.08 × 10³ km² out of 218.64 × 10³ km²). Together, these enrichments demonstrate that the zoning is interpretable not only via indicator contrasts, but also through recognizable land cover signatures, which “lock in” core protection and restoration targets while cleanly delineating development-carrying space.

Zone-level composition further clarifies the natural–human use gradient. EI and PR are forest-dominant (68.51% and 60.45% of zone area, respectively) with negligible built-up fractions (0.22% and 0.26%), whereas ID exhibits markedly higher cropland and built-up shares (27.01% and 3.21%). Derived indicators of natural cover and human utilization intensity reinforce this gradient. EI and PR maintain high natural cover (89.09% and 88.19%) and low human utilization (cropland + built-up: 10.91% and 11.80%), with high forest-to-cropland ratios (6.86 and 5.84), consistent with intervention and restoration spaces where ecological integrity and sink value remain dominant. LC functions as the long-term ecological backbone, retaining high connectivity-supporting cover (natural cover 83.90%) and a moderate utilization level (16.13%), consistent with sustained conservation under limited compatible use. RC occupies an intermediate regime (human utilization 22.73%), aligning with a management logic centered on risk-focused control in mixed mosaics. ID shows the lowest natural cover (69.79%), the highest human utilization intensity (30.22%), and the lowest forest-to-cropland ratio (1.53), consistent with concentrated development pressure. Notably, ID also exhibits the smallest mean effective unit area (27.40 km²), suggesting that development-carrying space is more fragmented and/or more strongly truncated by masking and boundary constraints, which helps explain why grid share can exceed area share for this zone. Figure 8 summarizes both the zonal composition (grid counts) and the resulting spatial configuration across Yunnan.

Cross-module validation was further corroborated by spatial overlay analysis with the corridor efficiency diagnostics (Figure 8). Emergency intervention (EI) cells exhibit systematic co-location with neighborhoods of high-connectivity corridors and areas proximal to ecological bottlenecks. This pattern indicates that the EI zone captures a composite condition defined by high ecological risk, high carbon sink value, and critical connectivity exposure, rather than merely serving as a proxy for any single indicator. An independent visual audit based on stratified random sampling of very-high-resolution imagery confirms these zoning semantics (Figure 9). EI sample locations are frequently associated with disturbed forest edges adjacent to roads and settlements, narrow ecological passages, or fragmented sensitive patches. In contrast, long-term conservation (LC) samples predominantly represent large, continuous natural cover within mountainous protected landscapes, while integrated development (ID) samples exhibit an urban–peri-urban mosaic structured by transportation infrastructure. The visual separability between risk control (RC) and priority restoration (PR) zones is comparatively weaker, consistent with their shared occurrence in mixed landscape mosaics. This observation suggests that enhancing the discrimination between risk control and restoration leverage will likely require the integration of finer disturbance proxies, such as edge pressure, road density exposure, and recent land change signals, alongside explicit spatial–structural descriptors to better represent heterogeneity.

Evidence source contributions reveal a structured division of labor in zone identification (

Table 4. Evidence contributions and robustness of fusion-based zoning.

Panel A. Evidence source contributions by zone
Zone	Surrogate Prob.	KG Prob.	Mechanism Prob.	Fused Prob.
EI	0.232	0.105	0.156	0.214
RC	0.184	0.199	0.190	0.204
PR	0.152	0.225	0.204	0.162
ID	0.284	0.175	0.182	0.281
LC	0.148	0.295	0.268	0.138
Panel B. Sensitivity analysis under different weight configurations
Weight_Surrogate	Weight_KG	Weight_Mechanism	Accuracy	Balanced Accuracy
0.60	0.20	0.20	0.643	0.712
0.50	0.20	0.30	0.653	0.711
0.50	0.30	0.20	0.649	0.711
0.40	0.40	0.20	0.646	0.703
0.40	0.30	0.30	0.644	0.702
0.30	0.50	0.20	0.630	0.688

, Panel A). Surrogate probabilities dominate the delineation of development-oriented (ID) and intervention (EI) zones, whereas KG-based semantic scores and mechanism-derived priors substantially enhance conservation-oriented zones—especially long-term conservation (LC). LC exhibits the highest contributions from both KG (0.295) and mechanism components (0.268), indicating that its delineation is governed by ecological structure and rule-consistent semantics rather than purely data-driven evidence. In contrast, ID and EI are more strongly influenced by surrogate probabilities, reflecting their association with heterogeneous development pressure and risk conditions. The fused solution achieves robust performance under the baseline weighting scheme (Panel B), with an overall accuracy of 0.649 and balanced accuracy of 0.711. Across alternative weight configurations, balanced accuracy remains stable (0.688–0.712), demonstrating that the zoning is not sensitive to parameter settings or dominated by any single evidence source.

Operationally, the zoning translates directly into differentiated planning actions. EI and PR constitute the highest-leverage intervention spaces, prioritizing immediate risk mitigation, avoided-loss protection, and targeted restoration where carbon and connectivity gains are most responsive. LC provides the structural backbone for long-term carbon sink stability and corridor maintenance. RC requires sustained monitoring and risk-centered management to prevent deterioration and spillover impacts. Crucially, ID, despite being the primary development-carrying space, must explicitly incorporate ecological risk constraints and corridor safeguards into development boundary delineation and intensity controls, because a large share of built-up land and substantial cropland pressure are concentrated within this zone. Figure 10 synthesizes the end-to-end KG-enabled zoning workflow into an implementable planning framework.

5. Discussion

5.1. From Productivity-Based Priority to Rule-Feasible Governance Decisions

This study advances carbon-oriented territorial governance by turning a productivity-informed priority surface into an implementable zoning decision under explicit feasibility rules [36]. The NPP-driven carbon sink priority (CSP) provides a planning-relevant representation of sequestration leverage, capturing where protection or restoration is expected to yield relatively larger potential carbon gains without conflating potential with net ecosystem balance. When embedded in the PRP indicator system and standardized alongside ESPC, ESI, and GDP intensity, CSP enables a coherent conservation–development gradient that remains consistent across modules: high-priority, vegetated landscapes dominate the upper tail of the prioritization ranking, while built-up and constrained landscapes accumulate in the low-priority tail.

Beyond this structural consistency, the framework explicitly links priority surfaces to rule-feasible decisions through knowledge graph constraints and probabilistic decoding, allowing ecological priorities to be translated into enforceable zoning allocations. At the same time, the results indicate that this translation is not deterministic: overlapping regimes and moderate classification accuracy (about 0.65) suggest that priority surfaces alone are insufficient to uniquely determine zoning outcomes. Instead, zoning emerges from the joint interaction of data evidence, governance rules, and mechanism constraints, highlighting the need to interpret zoning results as structured but non-unique solutions rather than exact partitions.

5.2. Independent Land Use Enrichment Validates Semantics and Exposes Actionable Leverage

Land use composition and enrichment patterns independently corroborate the governance logic. Built-up land is strongly concentrated in ID despite its smaller area share, while forests—the dominant carbon carriers—are preferentially secured within EI, PR, and LC. These signatures are consistent with the updated zone profiles. LC functions as the most stable ecological core, combining very high carbon potential with the lowest ecological risk. PR captures restoration-oriented landscapes where elevated ecological risk coincides with high carbon leverage. RC represents relatively stable maintenance space with positive connectivity and carbon potential, whereas EI reflects ecologically weak areas under strong human pressure. ID, in turn, represents mixed development landscapes under ecological constraints, where built-up land and cropland are disproportionately concentrated and development control must therefore internalize corridor and risk safeguards [37].

Operationally, this triad implies a differentiated action portfolio that is directly supported by the mapped outputs. EI and PR concentrate high-leverage interventions (rapid risk mitigation and restoration where carbon and connectivity gains are responsive), LC anchors long-term connectivity maintenance, and ID requires explicit integration of risk thresholds and corridor safeguards into development boundary delineation and intensity control. In practical planning contexts, this can be translated into concrete instruments such as zoning-based development quotas, ecological redline enforcement, restoration prioritization programs, and corridor protection regulations [38].

5.3. Limitations and Future Directions

Despite its contributions, the proposed PRP–KG–Bayesian zoning framework has several limitations that delineate clear priorities for subsequent research. First, the current risk characterization may inadequately represent fine-scale and rapidly evolving threats. The Ecological Risk Index (ESI) is computed on a 6 km × 6 km fishnet and subsequently interpolated to the planning grid, a procedure that inevitably smooths sub-kilometer landscape heterogeneity, particularly within complex urban–rural areas. Furthermore, the risk assessment remains largely static, relying on landscape composition and land use vulnerability scores without explicitly incorporating dynamic or episodic hazards [39]. Addressing these shortcomings will require integrating multi-scale risk diagnostics and leveraging remote sensing time series, including vegetation health anomalies, breakpoint detection, and impervious-surface expansion rates, to develop a dynamic monitoring and early-warning module capable of capturing both gradual degradation and sudden perturbations [40,41].

Second, while the Bayesian structural equation model (SEM) provides a calibrated mechanism narrative linking ecological structure, risk, and carbon potential, its current role is primarily to supply prior guidance rather than being jointly estimated with the zoning decision. Future work should explore tighter integration, for example by embedding SEM-derived latent factors or path coefficients directly into the multinomial model, or by specifying a unified hierarchical Bayesian framework that jointly estimates ecological processes and zoning decisions under shared uncertainty [42].

Third, although the knowledge graph rule base is essential for ensuring auditability and regulatory compliance, it remains partly dependent on expert-defined thresholds, which may introduce subjectivity and inter-zone overlap, particularly between the risk control and priority restoration zones in mixed landscape mosaics. In addition, the feasibility of zoning outcomes is sensitive to indicator distributions and data quality, implying that uncertainties in input datasets can propagate into final allocations. Future work should therefore focus on calibrating rules using empirical distributions, exploring probabilistic or fuzzy rule representations, and implementing versioned rule management. Beyond this structural consistency, the framework explicitly links priority surfaces to rule-feasible decisions through knowledge graph constraints and probabilistic decoding, allowing ecological priorities to be translated into enforceable zoning allocations. Priority surfaces alone are insufficient to uniquely determine zoning outcomes. Instead, zoning emerges from the joint interaction of data evidence, governance rules, and mechanism constraints, highlighting the need to interpret zoning results as structured but non-unique solutions rather than exact partitions.

6. Conclusions

This study develops a knowledge-guided probabilistic zoning framework that integrates pattern–risk–potential (PRP) indicators, knowledge graph rule encoding, mechanism-informed Bayesian calibration, and constrained posterior decoding. Applied to Yunnan Province, the framework translates ecological evidence, governance constraints, and probabilistic support into auditable territorial zoning outcomes under competing ecological and development objectives.

Three main findings emerge. First, the carbon sink priority surface reveals a clear conservation–development gradient, with high-priority areas concentrated in western and southwestern Yunnan and low-priority areas around major urban centers. Second, corridor diagnostics identify a central resistance belt and urban–rural bottlenecks, indicating that connectivity constraints are spatially concentrated in a limited number of structurally critical links rather than being uniformly distributed across the province. Third, the final zoning partitions the 11,853 effective planning cells into five governance regimes with distinct indicator profiles, land use signatures, and evidence contributions. Integrated development occupies 35.4% of grids and 30.16% of effective area, yet contains 78.85% of built-up land and 45.93% of cropland, whereas emergency intervention, priority restoration, and long-term conservation together secure 70.01% of forest area. Zone separability is generally strong, with one-versus-rest AUC values ranging from 0.7771 to 0.9950, although overlap remains greater among human-dominated governance regimes.

Methodologically, this work contributes an integrated analytical framework that combines Bayesian mechanism interpretation, knowledge graph rule encoding, and probabilistic zoning under explicit constraints. The results are derived from a single case study in Yunnan Province, and their transferability depends on the availability and quality of comparable datasets, as well as the compatibility of governance rules with local institutional contexts. In particular, the current implementation relies on indicator construction choices, spatial resolution, and expert-informed KG rules, which may introduce uncertainties when applied to different regions.

Overall, the proposed PRP-KG-Bayesian pipeline provides a transferable approach for carbon-oriented territorial zoning under competing ecological and development objectives. Future work should test the framework in diverse geographic settings, improve the integration of dynamic and multi-scale data sources, and further couple mechanism inference with zoning decisions to enhance process consistency.