Study area
The Hakka Cultural Ecological Protection Zones (CEPZs) system comprises three nationally designated protection zones spanning the mountainous borderland of Jiangxi, Fujian, and Guangdong provinces in South China (23°23′–27°08′ N, 113°50′–116°44′ E) (Figure 1A–D). The three zones — Ganzhou CEPZ in southern Jiangxi, Minxi CEPZ in western Fujian, and Meizhou CEPZ in eastern Guangdong — jointly cover 74,547 km2 and encompass 34 county-level administrative units (33 counties plus one municipal-district seat), forming the geographic core of the Hakka cultural sphere. Ganzhou CEPZ is the largest of the three (39,341 km2), containing 18 counties across the Ganjiang River headwaters and the Wuyi Mountain foothills; it hosts the highest concentration of Hakka enclosed dwellings (weilongwu) and the densest inland tulou distribution. Minxi CEPZ (19,353 km2) covers 6 counties centred on Longyan and Sanming, where the UNESCO-inscribed Fujian tulou clusters constitute the flagship built heritage. Meizhou CEPZ (15,853 km2) encompasses 9 counties on the middle reaches of the Meijiang River and is internationally recognized as the "Hakka Cultural Capital" with the highest per-capita overseas Hakka diaspora ratio.
The topography is dominated by mid-elevation mountains (400–1,600 m) belonging to the Wuyi, Nanling, and Lianhuashan ranges, with a northeast–southwest tectonic grain. The three zones share a subtropical humid monsoon climate: annual precipitation ranges from 1,500 to 2,100 mm, and the mean annual temperature is 18–21 °C. Broadleaf and mixed evergreen forests dominate the natural vegetation, interspersed with terraced cropland along the fluvial valleys. The three zones jointly house over 12 million people (2020 census) — a paradoxical combination of high heritage density and severe rural depopulation, with net out-migration exceeding 30% of registered residents in many hill counties. Hakka intangible cultural heritage (ICH) items registered at the national level number 23 across the three zones (Figure 1B–D), spanning performing arts (mountain songs, Hakka opera), traditional crafts (tulou construction, woodblock printing), and folk practices (San Yuan festivals, ancestor worship). The coexistence of dense heritage patrimony, contracting rural population, and comparatively intact mountain forests renders the Hakka CEPZs a distinctive comparative gradient for coupled ecological–structural robustness of the mapped ICH inventory network analysis at a subnational scale32. Basic administrative, morphological, and heritage attributes of the three zones are summarized in Table 1.
The CEPZ program was inaugurated by the Ministry of Culture and Tourism (MCT) in 2007 with the objective of safeguarding coherent territorial units in which ecological integrity and intangible heritage are conserved as a coupled system33. All three Hakka zones were listed at the national-priority level between 2013 and 2019, and administrative coordination is exercised by the provincial cultural affairs bureaus of Jiangxi, Fujian, and Guangdong, respectively. Since 2020, restoration and rehabilitation planning within CEPZ boundaries has been subject to the National Territory Space Planning (NTSP) framework, which requires spatially explicit prioritization of ecological corridors and heritage buffer zones34. The Hakka case, therefore, combines an unambiguous administrative jurisdiction with a spatially heterogeneous stressor regime, and its analytical outputs are directly actionable within existing planning instruments. Recent CEPZ-scale evaluations have called for network-based diagnostics to replace the inventory-based indicators previously in use35, setting the direct policy backdrop against which the DEHN framework is developed. The Hakka landscape is finally distinguished by its extensive diaspora legacy: Meizhou alone accounts for over one third of the global Hakka diaspora, and remittance-driven land management has produced land-use trajectories markedly distinct from those of demographically stable Chinese mountain regions36. This social layer is not directly parameterized in the present multilayer model but is documented here as the mechanistic backdrop against which the ecological and heritage layers evolve.
Data
Descriptive statistics on county areas within the study area: mean = 2,193 km2 (range: 721–3,946 km2; median: 2,089 km2; SD: 687 km2; n = 34 counties). The mean county diameter (assuming circular shape) is approximately 53 km, which exceeds the 10 km inter-layer coupling radius by a factor of 5.3. This systematic geocoding error means that the true ICH-ecological patch coupling could differ substantially from the centroid-based estimate. A sensitivity analysis increasing the coupling radius to 20 km showed that the top-20 RPI patch identity was preserved in 15 of 20 cases, suggesting moderate robustness to geocoding uncertainty. Village-level field surveys are identified as essential future work to resolve this limitation.
Table 2 summarizes the primary datasets used in this study. Land cover was derived from the China Land Cover Dataset (CLCD) developed by Wuhan University at 30 m spatial resolution, spanning 1985–2023 with annual increments37. Six representative years (2000, 2005, 2010, 2015, 2020, 2023) were retained to characterize multi-decadal change trajectories at consistent five-year intervals plus the terminal year. The CLCD schema distinguishes cropland, forest, shrub, grassland, water, ice/snow, and impervious surfaces, and its accuracy has been independently validated at an overall accuracy exceeding 79% over the study region38. Administrative boundaries and CEPZ perimeters were obtained from the Ministry of Culture and Tourism (MCT) national CEPZ registry and Gaode POI services; national-level ICH items were geocoded to the county centroid of their originating cultural custodian, following the convention used in prior Chinese ICH-network studies39. The composite dataset is released under CC-BY license and can be reproduced entirely through open remote-sensing archives, in keeping with recent calls for reproducible ecological-network research40.
Data pre-processing followed a five-step chain implemented in Python 3.11 with rasterio 1.3, GeoPandas 0.14, and NetworkX 3.2. First, the CLCD 30 m annual GeoTIFFs were subset to the three-zone union bounding box (23°23′–27°08′ N, 113°50′–116°44′ E) and reprojected to the Albers Conic Equal Area projection (lon₀ = 105°E, φ₁ = 25°N, φ₂ = 47°N) to preserve area for subsequent morphological analysis. Second, the union of the three CEPZ perimeters was rasterized as a study mask, and all off-mask cells were set to NoData throughout the pipeline. Third, cell counts by land-cover class were tabulated for each of the six benchmark years to support direct cross-year fragmentation-trajectory comparison. Fourth, the ICH point set was compiled from the State Council national-list registry (batches 1–5), geocoded to the county centroid of the declared cultural custodian, verified against publicly available point-of-interest services, and stored as a WGS-84 GeoJSON layer with attributes for item identifier, category (performing arts, traditional craft, folk practice), listing batch, and CEPZ affiliation. Fifth, all downstream vector–raster operations were carried out in Albers Conic Equal Area using windowed raster reads and vectorized in-memory array processing to preserve computational efficiency at the 30 m grid. All boundary and ICH source files, together with the reproducible pre-processing scripts, are available upon reasonable request.
Methods
The analytical chain (Figure 2) is organized as five horizontal swim-lanes — DATA, LAYER, COUPLING, DIAGNOSTICS, OUTPUT — and comprises six methodological modules: (i) morphological quantification of the ecological layer via a lightweight morphological spatial pattern analysis (MSPA-lite) on 30 m CLCD; (ii) spatial quantification of the heritage layer via kernel density estimation (KDE) and combinatorial adjacency graphs on the 23 national-level ICH items; (iii) coupling of the two layers into a bi-layer supra-network under a distance-decay inter-layer scheme; (iv) percolation-based resilience threshold identification under random and targeted node-removal rules applied to each layer independently; (v) a composite Restoration Priority Index (RPI) mapped over the ecological node set to identify Tier 1 (top-ranked) patches and top-priority corridors; and (vi) scenario simulation and multi-parameter sensitivity analysis on the priority typology and the coupling parameters.
The supra-adjacency matrix A (256 × 256) was constructed as a block matrix , where AE,norm and AH,norm are the intra-layer adjacency matrices normalised by their respective mean edge weights, and Ainter is the inter-layer coupling matrix. The matrix is symmetric (verified computationally: ||A - AT || < 1e-10) and contains no self-loops (trace(A) = 0).

Edge weight statistics before normalisation: ecological layer — min = 0.008730, mean = 0.098589, max = 1.618909; heritage layer — min = 0.006862, mean = 0.019848, max = 0.085832. After mean-normalisation: ecological — min = 0.0885, mean = 1.000, max = 16.4207; heritage — min = 0.3457, mean = 1.000, max = 4.3244.
Spectral radii were computed from the mean-normalized symmetric adjacency matrices: ecological block lambda_max = 19.6481, heritage block lambda_max = 10.5404, and full supra-network lambda_max = 19.6481. The ecological block, therefore, dominates the leading mode. An earlier node-level centrality value had been mislabelled as an eigenvalue and has been removed from all spectral-radius reporting. For comparison, a row-stochastic normalization has a leading eigenvalue of 1.000 by construction.
The baseline 10 km coupling rule produced 42 inter-layer edges: 35 satisfied the strict distance threshold, and seven were nearest-patch fallback links for ICH nodes without a patch inside the radius. Thus, all 23 ICH nodes retained at least one ecological connection. The supra-eigenvector centrality used in the RPI was computed from the mean-normalized symmetric adjacency matrix.
Ecological-layer quantification (MSPA-lite)
Morphological spatial pattern analysis (MSPA) partitions binary land-cover masks into topologically informative categories (core, edge, bridge, loop, islet, perforation, branch), thereby exposing habitat continuity independently of composition41. Because full MSPA on 30 m raster covering 74,547 km2 imposed prohibitive computational cost in preliminary trials, this study adopted a two-class MSPA-lite formulation that retains the core–edge distinction while collapsing bridge/loop/islet into an aggregated "edge" class. Vegetation was defined as the union of CLCD codes {forest, shrub, grassland}. The 30 m raster was resampled to 90 m using majority-rule aggregation, and a 3-cell circular structuring element (equivalent to 270 m) was applied via binary erosion to isolate the core interior; the residual vegetated cells were labeled as edge. Small patches (<5 km2) were excluded to focus on ecologically meaningful cores, following the size threshold widely adopted in Chinese regional MSPA studies42. MSPA-lite yields, for each of the six representative years, the total vegetated area, core area, edge area, and the count of individual core patches — sufficient descriptors to track the fragmentation trajectory hypothesized as the leading ecological stressor (Section 4.1).
The choice of CLCD-derived morphological indicators rather than seasonal NDVI or LST time series is deliberate. Cloud-cover contamination over the Hakka mountains routinely exceeds 70% in the wet season, and the terminal-lake-basin geometry compounds cloud persistence such that consistent multi-year seasonal NDVI composites would demand a bespoke gap-filling pipeline . Morphological indicators derived from annually validated categorical maps sidestep this atmospheric noise while preserving the connectivity information most relevant to network-based resilience analysis43.
MSPA-lite parameter sensitivity was probed in preliminary analysis. The core-erosion radius was varied across 2, 3, and 4 cells (equivalent to 180 m, 270 m, and 360 m in interior area at 90 m aggregation), and the minimum core area threshold was tested at 3 km2, 5 km2, and 10 km2. The final parameterization (3-cell erosion, 5 km2 threshold) was retained because it preserved a stable rank order of patch abundance across the six years while eliminating spurious small cores generated by CLCD classification noise. Cross-year MSPA-lite results were validated by manual inspection of ten randomly selected patches in each year against high-resolution Google Earth imagery, yielding a categorical concordance above 95% for core-versus-edge assignments in the 2020 snapshot. Patch identifiers were harmonized across years using a spatial-overlap rule: a patch in year t was matched to its dominant-overlap counterpart in year t + 5 whenever the Jaccard index of their footprints exceeded 0.60. Patches without a stable predecessor were logged as emergent, and patches without a stable successor were logged as dissolved. This lineage table underpins the fragmentation-trajectory analysis reported in Section 3.1.
Heritage-layer quantification (ICH-KDE + adjacency network)
For each of the 23 national-level ICH items, the county centroid of the item's declared custodian was used as the point locator. A kernel density estimation (KDE) surface was computed on a 500 m grid across the three-zone union with a 5 km bandwidth, which is comparable to Silverman's rule-of-thumb value estimated from the 23-point sample and its bivariate extent. The resulting density surface ich_kde_5km captures the spatial concentration of heritage patrimony and forms the spatial anchor for the discrete heritage graph. Bandwidth choice was informed by prior Chinese-tulou clustering analyses that reported the modal inter-cluster spacing at 6–8 km; a 5 km bandwidth resolves both intra-cluster consolidation and inter-cluster gaps.
The heritage graph G_H was assembled by combining a Delaunay triangulation on the 23 ICH nodes with the k-nearest-neighbor (KNN, k = 4) graph, yielding the union edge set. This combinatorial approach eliminates the elongated Delaunay edges spanning topographic barriers while retaining nearest-neighbor connectivity, following the graph construction protocol adopted in recent multiplex ecosystem-service studies44. Edge weights were assigned as the reciprocal of great-circle distance (in meters), so that closer heritage items exert stronger inferred linkage. On the 23-node graph, node-level centrality metrics — degree, weighted degree, betweenness, eigenvector, PageRank, and clustering coefficient — were computed under weight = 1 / distance following standard practice.
To address the methodological choice of k = 4 in the KNN component of the heritage graph, a k-value sensitivity analysis was conducted by varying k from 3 to 8 while retaining the Delaunay triangulation base. The consensus percolation threshold ranged from 0.754 (k = 4) to 0.923 (k = 7), with intermediate values of 0.779 (k = 3), 0.773 (k = 5), 0.852 (k = 6), and 0.885 (k = 8). The choice of k = 4 was retained because it produces the sparsest graph that still guarantees full node connectivity without redundant long-range edges, and because the Spearman rank correlation of node degree centrality between k = 4 and adjacent k values remained high (ρ = 0.691 for k = 3, ρ = 0.793 for k = 5). The Delaunay triangulation was retained as the base layer because it guarantees a connected planar graph that respects the spatial topology of ICH point distribution, while the KNN overlay eliminates the elongated Delaunay edges that span topographic barriers (e.g., Wuyi mountain ridge). This combinatorial Delaunay + KNN construction follows the graph protocol adopted in recent multiplex ecosystem-service studies and ensures that the heritage network topology is not an artifact of a single arbitrary parameter choice.
Ecological corridors and least-cost paths
Resistance surface construction followed the class-based look-up-table (LUT) convention45. Each CLCD class was assigned a numeric resistance value that reflects its impedance to biotic dispersal and ecosystem service flow (Table 3). Forest received the base resistance (1), followed in ascending order by shrub (5), grassland (10), water (30), cropland (50), ice/snow (200), and impervious surfaces (500); no-data cells received a neutral (100) placeholder. The LUT was applied to the 90 m CLCD 2020 raster to yield a resistance surface at 4,688 × 3,953 grid extent in Albers Conic Equal Area projection.
Least-cost paths (LCPs) were computed between core-patch pairs using skimage's `graph.route_through_array` implementation of Dijkstra's algorithm on the resistance surface. Candidate node pairs were restricted to the union of the K-nearest-neighbor (k = 4) and Delaunay-triangulation graphs of the 233 patch centroids in projected space, following the LCP-graph protocol widely used in Chinese regional connectivity studies46. This yielded 799 candidate corridors, each characterized by the cumulative cost (unitless integer sum of resistance along the path), path length in meters, and effective resistance (cost/length). All 799 corridors were retained in the final ecological graph G_E, since none exceeded a maximum-cost cut-off recommended for regional connectivity studies47.
Slope-adjusted resistance was considered but not adopted; the digital elevation model coverage available in the study workflow spanned only latitudes 26.00–27.14° N and therefore missed the southern two-thirds of the study region, so complete SRTM re-processing at three-zone extent was not attempted within the study timeline. A pure LULC resistance parameterization is a standard fallback in Chinese regional corridor studies where DEM completeness is not achievable, and it isolates the LULC signal without confounding topographic gradients48.
Implementation of the LCP computation employed skimage.graph.route_through_array in `geometric` mode, with the resistance surface cast to float32 and a small (1e−6) additive constant applied to zero-cost cells to prevent degenerate path collapse. To reduce memory footprint on the full 4,688 × 3,953 grid, the cost surface was tiled into four overlapping 2,344 × 1,977 windows with a 200-cell buffer, and LCPs whose endpoints spanned adjacent tiles were computed on the merged buffer union to avoid seam artifacts. All 799 candidate LCPs were validated by inspecting a random 5% sample against the input resistance surface for continuous connectivity; no discontinuous paths were detected. Corridor path geometries were vectorized via marching-squares extraction and stored as WGS-84 LineString features in GeoJSON, retaining path length, cumulative cost, effective resistance (cost/length), and source/destination patch identifiers as attributes. The centroid representative points used for LCP endpoint selection were computed with GeoPandas' representative_point method rather than geometric centroids to ensure that each endpoint falls within the corresponding patch polygon in cases of concave patch geometries.
Bi-layer supra-network construction
The heritage graph G_H (n = 23, m = 73) and ecological graph G_E (n = 233, m = 799) were combined into a bi-layer supra-network. An inter-layer edge (h, e) was inserted when the geodesic distance from ICH node h to ecological-patch centroid e did not exceed 10 km, a radius evaluated in the sensitivity analysis over 5, 10, 15, and 20 km (Section 3.5). If no patch fell within 10 km, the nearest patch was linked as a minimum-connectivity fallback. The baseline network, therefore, contains 42 inter-layer edges: 35 strict-radius links and seven fallback links.

Edge weights in the supra-adjacency matrix A (256 × 256) were assigned as follows: intra-heritage edges retained their reciprocal-distance weights; intra-ecological edges received the reciprocal of the least-cost path cost (1 / cost); and inter-layer edges received as defined below, where d is the coupling distance in kilometers, and
w_intra
is the mean intra-layer edge weight, yielding a smoothly decaying inter-layer coupling calibrated to the intra-layer magnitude. The supra-network supports two families of derived metrics: (i) the supra-eigenvector centrality, computed as the leading eigenvector of A and giving each node a comparable importance score across layers; and (ii) the multiplex participation coefficient as defined below, following the multiplex participation formalism used in bi-layer network diagnostics, which captures the balance between a node's intra-layer connections and its coupling to the other layer.


The supra-adjacency matrix A was stored as a sparse CSR matrix using SciPy's sparse module. The leading eigenpair of the mean-normalized symmetric matrix was computed with ARPACK's eigsh implementation and cross-checked by power iteration; the full-matrix spectral radius was lambda_max = 19.6481. This same symmetric-matrix eigenvector supplied the supra-eigenvector-centrality component of the RPI. Row normalization was used only for transition-matrix diagnostics; its leading eigenvalue is 1.000 by construction. Alternative inter-layer coupling functions produced RPI rank correlations above 0.94 with the exponential-decay baseline.
Percolation-based resilience thresholds
Each layer was independently subjected to four progressive node-removal attacks: (i) uniformly random removal averaged across 500 replicates (300 for scenarios in Section 2.3.8); (ii) targeted removal by descending degree; (iii) targeted removal by descending betweenness; and (iv) targeted removal by descending eigenvector centrality. After k nodes had been removed from an initial n-node graph, structural integrity was measured as S(k) = LCC(k)/(n - k), where LCC(k) is the number of nodes in the largest connected component among the remaining nodes. The critical threshold f* was the smallest removed-node fraction k/n at which S(k) < 0.5. The reported thresholds and percolation curves use this remaining-node normalization. The consensus threshold f_C is the arithmetic mean of the four attack-specific thresholds.
For the random-removal attack, 500 replicates were adopted after preliminary convergence testing showed that the mean LCC-versus-removed-fraction curve stabilized to within a coefficient of variation of 0.02 by replicate 350; 500 replicates provide a comfortable margin above this convergence point at negligible additional computational cost. Ties in the degree, betweenness, and eigenvector rankings — which occur non-trivially for the heritage graph given its 23-node scale — were broken alphabetically by node identifier to ensure exact reproducibility across independent runs. Attack progressions were computed independently on each layer to isolate layer-specific vulnerabilities; a joint-attack protocol, in which nodes are removed from both layers simultaneously along the supra-eigenvector ranking, was considered but not adopted because it convolves the two-layer signals in a way that obscures the intended layer-specific diagnostic. The 0.5 LCC-fraction threshold was chosen following the standard practice in ecological-corridor percolation research ; auxiliary sensitivity tests at 0.4 and 0.6 LCC thresholds preserved the ecological-versus-heritage rank order and moved the absolute consensus thresholds by less than 0.05 in either direction. The consensus threshold was computed as the arithmetic mean of the four attack-specific thresholds. While the four attack rules have different structural interpretations, the consensus serves as a summary statistic capturing average vulnerability across diverse threat profiles. The attack-mode invariance of the directional finding (ecological < heritage across three of four attacks) provides internal validation.
For random attacks, 500 independent replicates were conducted. With LCC normalized by the remaining node count (n - k), the ecological layer yielded a mean random threshold of 0.623 ± 0.058 (SD), and the heritage-inventory layer yielded 0.960 ± 0.082. Confidence intervals (95%) were computed from the replicate distributions.

Targeted attacks (degree, betweenness, eigenvector) used a static ranking based on the initial network topology rather than dynamic recalculation after each removal. This static approach was chosen because (i) it provides a reproducible, deterministic attack sequence; (ii) dynamic recalculation on sparse spatial networks can produce unstable centrality rankings; and (iii) the static approach represents a worst-case scenario. Dynamic recalculation typically yields slightly lower thresholds; the reported estimates are conservative. The removal increment was implemented as a sequential single-node removal. For the 23-node heritage network, each removal corresponds to ~4.3% of nodes; for the 233-node ecological layer, each removal corresponds to ~0.43%. This finer-than-0.025 resolution ensures accurate threshold detection.
Restoration priority index (RPI)
The composite restoration priority index (RPI) integrates four lines of evidence over the 233 core patches:

where z(·) denotes standardization to zero mean and unit variance across all patches, w1 = 0.35 emphasizes bi-layer structural centrality, w2 = 0.20 assigns higher priority to small patches (fragmentation hot spots), w3 = 0.30 promotes patches with strong ICH coupling, and w4 = 0.15 up-weights isolated patches with high mean edge cost. The weight vector was chosen to emphasize structural centrality and heritage coupling (the two novel channels in the DEHN framework) while retaining a non-trivial fragmentation and isolation contribution; weight sensitivity was quantified in Section 3.5. Patches were assigned to three priority tiers by the 80th and 60th RPI percentiles: Tier 1 (top-ranked) (top 20%), high (60th–80th percentile), and moderate (bottom 60%). Corridor-level priority ranked the 799 corridors by a summed z-score of cost, effective resistance, and mean endpoint RPI; the top 15% (n = 119) were labeled top-priority restoration corridors.
Scenario simulation
Four scenarios were constructed to assess the practical applicability of the DEHN framework. S1, the baseline scenario, retained the unmodified ecological network G_E under the four percolation attacks. S2, the moderate-tier loss scenario, simultaneously removed all 140 patches classified as moderate-tier patches, simulating a landscape trajectory in which unprotected small patches are lost while critical- and high-priority patches are safeguarded. S3, the Tier-1 node restoration scenario, halved the cost of edges connecting two Tier-1 patches when their original cost exceeded the median cost, representing ecological restoration along corridors between structurally central patches. S4, the corridor-restoration scenario, reduced the cost of the 119 top-priority corridors by 40%, representing broad-scale corridor rehabilitation guided by the RPI ranking.
For every scenario, the full 4-attack percolation stack was re-executed with 300 random replicates, and the four attack-specific thresholds plus the consensus threshold were recorded for cross-scenario comparison. Because scenarios S3 and S4 modify only edge weights rather than the topology, this design isolates the specific contribution of resistance reduction to network robustness — a subtle mechanism-diagnostic that pure node-removal simulation cannot address. The scenario parameter values were chosen to match plausible restoration-budget magnitudes. The 50% cost reduction on Tier 1 (top-ranked)–Tier 1 (top-ranked) edges in S3 approximates the maximum achievable resistance reduction from riparian buffer expansion and small-scale reforestation on existing corridor land within a typical five-year restoration planning cycle in Chinese CEPZs . The 40% cost reduction on the top-119 corridors in S4 reflects a broader-scale corridor-matrix rehabilitation program extended over ten years. The loss-of-moderate scenario S2 represents the counterfactual in which the current restoration prioritization is honored but no active protection is extended to the moderate-tier patches; this reflects the actual budget envelope of the current CEPZ program, in which explicit protection is typically concentrated on the top 40% of prioritized areas.
Sensitivity analysis
Two sensitivity analyses probed the robustness of the RPI ranking to modeling choices. First, each RPI weight (w1 – w4) was perturbed by ±0.05 and ±0.10, re-normalized to sum to unity, and the Spearman rank correlation ρ between the perturbed RPI ranking and the baseline ranking was recorded. Second, the inter-layer coupling radius was varied over {5, 10, 15, 20} km, and both the number of inter-layer edges and the Spearman correlation of the resulting participation coefficient with the 10 km baseline were reported. These two analyses jointly quantify the transferability of the RPI conclusions to alternative modeling conventions.
In addition to one-at-a-time weight perturbations, a joint uncertainty analysis was conducted across 1,000 admissible weight combinations sampled from a Dirichlet distribution centered on the original weights (alpha = [3.5, 2.0, 2.5, 2.0]). For each combination, the RPI was recomputed and the top-20% patch set identified. Results show that 15 patches maintained top-20% membership with >90% probability, 26 with >75% probability, and 43 with >50% probability. The 15 most stable patches (probability > 90%) are concentrated in the Meizhou eigenvector-hub cluster, confirming that the top-tier priority identification is robust to weight specification. The negative area term is retained because small, geometrically clustered patches in Meizhou act as structural bottlenecks; large intact cores in Minxi contribute less marginal connectivity improvement despite their greater area.
Weight perturbations of ±0.05 and ±0.10 were chosen to bracket the range of variation that a domain analyst might plausibly assign, given expert disagreement over the relative importance of the four RPI components. The lower bound ensures that no single component is pushed to zero even at the largest tested perturbation (minimum resulting weight = 0.05), preserving all four evidence lines in every perturbation. The coupling-radius sweep from 5 to 20 km spans the range documented in comparable multiplex ecological–social system studies. Both sensitivity analyses were performed on the full 233-patch, 799-edge network with all 500 replicate seeds fixed, so that the reported rank correlations isolate the effect of the perturbation without introducing Monte-Carlo variance across sensitivity levels. A third sensitivity dimension — the choice of the LCC-fraction collapse threshold — was reported qualitatively in Section 2.3.5 and further discussed in Section 4.4 alongside the framework's other bounded limitations.