Rural Settlement Turnover Across Cross-Administrative Geographical Contexts in a Plateau–Mountain Region: Evidence from Yunnan, China

Abstract

Administrative boundaries and net area change can obscure rural settlement dynamics in geographically heterogeneous mountain regions. We examined these two sources of information loss in Yunnan, China, using a state–balance–process–interface framework. Settlement persistence, gain, and loss were mapped from 30 m land use data for 1990, 2000, 2010, and 2020. Five cross-administrative geographical context units (GCUs) were delineated on a 1 km grid, and turnover occurrence and conditional positive intensity were compared on a 10 km grid. Settlement area increased by 384.29 km² over 1990–2020, whereas cumulative turnover reached 1690.36 km², or 4.40 times the net increase. Conservative filtering of immediate reversals reduced turnover to 1117.16 km², or 2.91 times the net increase. Both estimates retained the ordering 2010–2020 > 2000–2010 > 1990–2000. However, filtering retained only 9.46% of early-period turnover, which made its absolute magnitude particularly sensitive to temporal consistency. In 2010–2020, a near-zero net change (−20.09 km²) coexisted with In 2010–2020, near-zero net change (−20.09 km²) coexisted with substantial turnover under both baseline and conservative accounting (916.28 and 699.56 km², respectively). GCU rankings differed across settlement state, net balance, and turnover. Lake–Basin had the highest density and turnover intensity, Hill–Mountain Transition contributed the largest absolute net gain, and Alpine Canyon had almost no turnover. Cropland supplied 71.88% of cumulative settlement gains, while other built-up land became an increasingly prominent destination of settlement losses. Independent geographical references and spatially blocked administrative comparisons supported the GCUs’ geographical relevance and modest additional information. The framework identifies contrasting land transitions for regional monitoring and planning assessment. Its wider application requires locally defined contexts and independent testing; it does not establish policy effects or ecosystem condition.

1. Introduction

Rural settlements connect residential and productive activities with the land systems that support them. In mountain and peripheral regions, urbanization, migration, agricultural change, infrastructure development, depopulation, and land abandonment can occur simultaneously in neighboring places. Elevation, relief, karst hydrology, basin lowlands, and accessibility further differentiate these trajectories. Research in the Colombian Andes links sustainability pathways to environmental gradients, remoteness, and competition for land [1]. Evidence from Japan also shows that biodiversity responses to human depopulation vary among settings [2]. Regional analysis must therefore explain how rural settlement land changes while retaining the geographical differences that shape those changes.

Research on rural settlements increasingly examines spatial restructuring, village life cycles, land use transitions, and sustainable rural development [3,4,5,6,7]. Mountain studies also consider topographic gradients, livelihood trajectories, and relationships across spatial scales [8,9]. Other work has examined changes in human settlement quality and the ecological consequences of settlement change [10,11]. These studies link rural settlement dynamics to interacting environmental, locational, and socioeconomic conditions. Yet regional comparisons remain vulnerable to two forms of information loss. Administrative averages can combine contrasting geographical settings, while endpoint area and net change can conceal simultaneous gains and losses. Resolving both problems requires attention to the spatial units of comparison and the processes represented by change metrics.

The first problem concerns the spatial units used to describe change. Counties and townships organize statistical reporting, public-resource allocation, and planning, but their boundaries may divide continuous ridge–valley systems, karst landscapes, lake basins, or border corridors. Such mismatches between environmental processes and governance boundaries can alter relationships observed after aggregation [12]. Urban–rural gradients capture continuous variation, but individual gradients may not represent the combinations of constraints encountered in mountain regions [13]. We use geographical context units (GCUs) as an intermediate analytical layer between pixels and administrative jurisdictions. These mutually exclusive units group areas with similar topographic, hydrological, karst, and locational conditions. They allow us to test whether geographical context adds information to administrative categories while retaining those categories’ institutional role.

The second problem concerns the processes concealed by net change. Identical net gains can result from limited expansion or from extensive, offsetting gains and losses. Global land use research shows that gross change can substantially exceed long-term net change when repeated and bidirectional transitions are counted [14]. Time-series decompositions likewise distinguish quantity, exchange, and spatial shift from endpoint differences [15]. We define turnover as the sum of settlement gain and loss areas in each adjacent interval. Summing these events measures conversion activity, including repeated changes at the same location. It does not measure the unique area changed or identify village relocation, homestead reclamation, or urban absorption. Because many cells record no turnover, we distinguish its occurrence from its intensity among cells with positive values. Source–destination composition then identifies the land use categories involved in these transitions.

Yunnan provides a heterogeneous plateau–mountain setting in which both problems can be examined. Lake basins, karst mountains, hill–mountain transitions, alpine canyons, and border corridors occur in close proximity. Studies of urbanization, cross-border transport, the Erhai Lake Basin, and karst land use transitions document distinct constraints and trajectories across these settings [16,17,18,19]. Their biophysical and locational combinations frequently extend across county and prefectural boundaries. Yunnan therefore offers a useful test of whether geographical context complements administrative comparisons and whether net change adequately describes conversion activity. Its geographical specificity also means that local thresholds and class definitions require reassessment before application elsewhere.

We examine how spatial support and change metrics affect the regional diagnosis of rural settlement dynamics. Three questions guide the analysis. First, can knowledge-guided rules produce reproducible GCUs that agree with independent geographical references and add information beyond administrative categories? Second, how do settlement stock, net balance, turnover, and source–destination composition vary among GCUs, and how do zeros and initial settlement stock affect those comparisons? Third, which findings persist across alternative spatial supports, classification rules, temporal-consistency diagnostics, and spatial-dependence checks? We address these questions through a GCU-based state–balance–process–interface framework. The intended application is regional screening for differentiated planning, with explicit limits on causal interpretation and transferability.

2. Materials and Methods

2.1. Study Area

Yunnan lies in southwestern China (21°08′–29°15′ N, 97°31′–106°11′ E), between the southeastern Tibetan Plateau, the Yunnan–Guizhou Plateau, and the low-latitude mountains of Southeast Asia. Mountains, plateaus, and deeply incised valleys dominate the province. Lakes, basins, karst landscapes, and the alpine canyons of the Hengduan Mountains create pronounced topographic and hydrological contrasts (Figure 1). High mountains and deep valleys characterize western and northwestern Yunnan, whereas karst landscapes are widespread in the east and southeast. Central lake basins and broad valley floors concentrate agricultural and construction land. Studies elsewhere in southwest China and the Himalayan region also link rural settlement patterns to topographic constraints and land availability [20,21,22].

Figure 1. Location and topographic setting of the study area.

Yunnan’s analytical value lies in the coexistence of contrasting geographical contexts within one province. A single jurisdiction can contain several contexts, while each context may extend across county or prefectural boundaries. We therefore compare rural settlement change across GCUs within a common provincial domain. This design tests whether administrative aggregation conceals geographical variation while retaining administrative units as the basis for governance.

2.2. Data and Spatial Supports

We used 30 m CNLUCC/LUCC land use data for 1990, 2000, 2010, and 2020 from the Resource and Environment Science and Data Center (RESDC), Chinese Academy of Sciences. Rural settlements corresponded to second-level land use code 52. All rasters were projected to WGS 84/UTM Zone 47 N and aligned with the 1990 raster in pixel size, extent, and grid origin. Retaining only pixels valid in all four years yielded a common valid area of 383,617.5291 km². Pixel-wise overlays identified settlement persistence, gain, and loss within this area. The data support regional land use transition analysis but cannot resolve individual buildings, homestead parcels, or changes within villages.

Geographical context and settlement change were represented at different spatial resolutions. Context inputs described elevation, slope, relief, rivers, lakes, basins, karst, rocky desertification, administrative boundaries, and border corridor conditions. Sources included NASA Earthdata, the National Platform for Common Geospatial Information Services, RESDC, and the National Earth System Science Data Center. These inputs were summarized on a 1 km grid to construct GCUs independently of observed settlement change. Exact area overlays linked the GCUs to a regular 10 km grid for statistical comparison and the principal maps. Thus, 30 m pixels recorded change, 1 km cells defined geographical context, and 10 km cells supported comparison and inference. The GCU domain contained 387,587 valid 1 km cells, whereas the common valid land use area covered 383,617.5291 km². These domains served different purposes and were kept distinct in area calculations and aggregation.

We assessed geographical concordance using reference layers that had not contributed to GCU delineation. These comprised the 1:1,000,000 geomorphological dataset of China [23] and the Global Lithological Map (GLiM) [24]. Period-matched Landsat true-color and near-infrared false-color imagery provided a separate visual check of representative settlement gains and losses (Figure S2). This check provided image-based quality control without estimating map-wide accuracy. We also applied an immediate-reversal filter to assess sensitivity to temporal inconsistency in the land use series. Such inconsistency may arise from differences in sensors, reference materials, or production workflows. Table 1 summarizes the datasets and their analytical roles.

Table 1. Data sources, spatial resolutions, and analytical roles.

Data	Dataset/Source	Resolution/Period	Main Analytical Use
Rural settlement and land use	CNLUCC/LUCC, RESDC, Chinese Academy of Sciences	1990, 2000, 2010, 2020; 30 m	Settlement stock, gain, loss, turnover, and source–destination transitions
GCU context inputs	DEM, NASA Earthdata; river/lake/basin and boundary data, National Platform for Common Geospatial Information Services; karst data, National Earth System Science Data Center; rocky desertification data, RESDC	30 m and vector data summarized to 1 km	Topographic, hydrological, karst, and border corridor context for GCU construction
Administrative units	Provincial, prefecture/county, and international boundaries, National Platform for Common Geospatial Information Services	Vector	Administrative comparison and border context
Independent geographical references	China 1:1,000,000 geomorphological dataset [23]; GLiM lithological map [24]	Independent reference data	External geographical concordance
Landsat imagery	Period-matched Landsat true- and false-color imagery	30 m	Representative visual cross-check of gain/loss cases
Analytical supports	Derived in this study	1 km GCU grid; 10 km statistical grid	GCU construction; KEEP20 N = 3987; DOM60 N = 3282

Note: External geomorphology and GLiM data were used only for geographical concordance and were not inputs to GCU construction.

2.3. Methods

The workflow combined GCU delineation, settlement-change accounting, two-part statistical analysis, and checks of reproducibility and robustness (Figure 2). Five mutually exclusive GCUs were constructed on the 1 km context grid. Settlement persistence, gain, and loss were then mapped on the common valid 30 m land use area, together with source–destination transitions. Aggregation to the 10 km grid enabled separate comparisons of turnover occurrence and conditional positive intensity. We tested sensitivity to spatial support, classification rules, temporal consistency, initial stock, and spatial dependence. Independent geographical references and administrative comparisons assessed the GCUs’ geographical concordance and additional information.

Figure 2. Workflow for GCU delineation, settlement-change accounting, and assessment of robustness, reproducibility, and geographical concordance.

2.3.1. GCU Delineation, Reproducibility, and Geographical Concordance

We defined geographical context through topography, hydrology, lake–basin setting, karst conditions, border location, and corridor structure. These attributes were represented by five mutually exclusive GCUs on the 1 km grid. The classes were Plateau Lake–Basin, Karst Mountain–Rocky Desertification Sensitive, Plateau Hill–Mountain Transition, NW Yunnan Alpine Canyon, and Border Mountain–Corridor. We refer to them as Lake–Basin, Karst, Transition, Alpine Canyon, and Border–Corridor, respectively. Classification combined geographical knowledge with observed data distributions and spatial continuity. The rules were designed for comparisons within Yunnan, rather than as universal thresholds for natural regionalization. Research on ecoregionalization, settlement–topography relationships, rocky desertification, and cross-border corridors informed variable selection and interpretation [25,26,27,28,29]. These studies did not supply the Yunnan-specific operational thresholds.

Overlapping classes were resolved in the fixed order of Alpine Canyon > Lake–Basin > Karst > Border–Corridor > Transition. The executable procedure used four fixed binary input layers: SAlpCanV6, LakeGateV8, SKarst5, and BorderGateV6. A value of 1 triggered assignment to the corresponding class according to that priority; unassigned cells entered Transition. Earlier Alpine Canyon criteria combined mean elevation ≥2300 m, mean slope ≥12°, mean relief ≥250 m, and river distance ≤8 km. Earlier Karst criteria used karst coverage ≥0.20 or rocky desertification coverage ≥0.10. Lake–Basin and Border–Corridor were assigned using the finalized LakeGateV8 and BorderGateV6 binary layers, respectively, under the fixed overlap-priority order. Earlier distance and combination rules documented their development and supported sensitivity tests but did not exactly reconstruct the final layers. Table 2 and Algorithm S1 specify the final inputs, priority order, and cell-wise reproduction procedure.

Table 2. Operational definitions and overlap priorities of the five geographical context units.

Code	GCU	Operational Definition	Priority
1	Plateau Lake–Basin	LakeGateV8 = 1; plateau lake/basin/flat-valley hydrological setting; 1 km cells = 15,905	2
2	Karst Mountain–Rocky Desertification Sensitive	SKarst5 = 1; karst-sensitive mountain context; 1 km cells = 54,806	3
3	Plateau Hill–Mountain Transition	Residual after SAlpCanV6, LakeGateV8, SKarst5 and BorderGateV6 assignments; 1 km cells = 224,817	5
4	NW Yunnan Alpine Canyon	SAlpCanV6 = 1; alpine canyon context; historical core criteria: DEM ≥2300 m, slope ≥12°, relief ≥250 m, river distance ≤8 km; 1 km cells = 30,567	1
5	Border Mountain–Corridor	BorderGateV6 = 1; border location and mountain corridor context; 1 km cells = 61,492	4

Note: Priority is used solely to resolve overlapping assignments and does not imply a developmental or ordinal hierarchy. SAlpCanV6, LakeGateV8, SKarst5, and BorderGateV6 are the finalized operational binary layers developed through iterative testing and refinement of alternative threshold combinations and spatial rules. This process was used to identify specifications that most consistently represented the intended terrain, hydrological, karst, and border-related geographical contexts while maintaining spatial coherence and classification stability across Yunnan. The resulting layers were subsequently retained as the fixed operational inputs for the final GCU classification. Exact reproduction therefore requires these finalized binary layers and the prescribed priority order.

Operational reproducibility and external geographical validity were evaluated separately. The final GCU classification was rerun from the four archived operational input layers under the fixed priority rules, and cell-by-cell agreement with the finalized map was assessed. Standardized profiles of the construction variables were used to characterize the internal structure of the classes. Independent external evaluation was then conducted using the China 1:1,000,000 geomorphological dataset and GLiM lithology, neither of which contributed to GCU construction. These references were used to examine whether the five GCUs exhibited independently observed geomorphological and lithological characteristics consistent with their geographical definitions. In the absence of an authoritative five-class regionalization for these composite contexts, this evidence was interpreted in terms of convergent geographical validity rather than conventional map-classification accuracy.

2.3.2. Settlement-Change Accounting and Source–Destination Analysis

Land use code 52 was converted to a binary settlement indicator within the common valid area. Adjacent-year rasters were overlaid for 1990–2000, 2000–2010, and 2010–2020. For interval a–b, persistence (P) denoted settlement in both years, gain (G) denoted conversion to settlement, and loss (L) denoted conversion from settlement. Net change (NC) and turnover (T) were calculated as follows.

$$NC(a,b)=G(a,b)-L(a,b); T(a,b)=G(a,b)+L(a,b)$$

We checked the stock-balance identity, end stock–start stock = G − L, for each interval. We also verified P + L = start stock, P + G = end stock, and T = G + L. This accounting follows process-oriented intensity analysis and quantity–exchange decomposition [30,31]. Mean annual turnover intensity was calculated as ATI = T/(S × Δt) × 100%, where S is the cell’s common valid area and Δt is interval length. ATI is expressed in % yr⁻¹. Cumulative gain, loss, and turnover sum the event areas across the three intervals. A pixel changing in several intervals contributes once to each event. Cumulative turnover therefore measures conversion-event area rather than the unique area changed between 1990 and 2020.

Source–destination analysis used overlays of the original land use codes. Gain sources were identified from the initial class, and loss destinations were identified from the final class. Non-settlement classes were grouped as cropland, forest, grassland, water, other built-up land, and unused land. Each source share used total settlement gain as its denominator; each destination share used total settlement loss. These compositions describe transitions between land use categories, without identifying reclamation, restoration, urban absorption, or other mechanisms.

Turnover can be more sensitive than net change to inconsistencies among maps. We therefore applied a conservative diagnostic that excluded immediate reversal events identified across consecutive intervals. Baseline and filtered estimates were compared for temporal ordering and GCU contrasts. The filter tests sensitivity to temporal consistency; it does not establish that every excluded reversal was a classification error.

2.3.3. Statistical Grid and Two-Part Spatial Inference

Analysis used a regular 10 × 10 km grid. We retained cells with common valid area covering at least 20% of their nominal area (KEEP20), yielding 3987 cells. Exact overlays assigned each cell to a GCU only when that class occupied at least 60% of its valid classified area (DOM60). This retained 3282 assigned cells for between-GCU comparisons and excluded 705 mixed cells. The assigned cells comprised the complete analytical population under the chosen extent and thresholds rather than a probability sample of mountain regions. We therefore interpreted p-values as diagnostic of differentiation within this defined Yunnan domain.

The large number of zero values motivated a two-part analysis. First, turnover occurrence was defined as O = I (T > 0). We compared occurrence proportions using Pearson’s χ² and Cramér’s V, with empirical p-values from 9999 permutations restricted to 50 km spatial blocks. GCU labels were shuffled within blocks while the occurrence pattern was held fixed. The empirical p-value equaled (1 + the number of permuted statistics at least as large as observed)/(9999 + 1). Second, conditional positive ATI was compared using the tie-corrected Kruskal–Wallis test [32]. The rank-based effect size was ε² = (H − k + 1)/(N − k), where H is the test statistic, k is the number of groups, and N is the positive-cell count. Significant omnibus tests were followed by Dunn pairwise comparisons with Holm adjustment [33,34]. This approach retained zero values as observations and distinguished occurrence from intensity conditional on occurrence.

Alpine Canyon had no positive cells in the first two periods and only two in 2010–2020. Its occurrence pattern was therefore interpretable, but the positive observations were insufficient for a substantive intensity ranking or mechanistic interpretation.

Global Moran’s I with Queen contiguity quantified spatial dependence. We assessed stability using 20 km random spatial thinning, a 50 km spatial block bootstrap, and 30 km thinning as a more restrictive stress test. Conley heteroskedasticity- and autocorrelation-consistent inference (HAC) provided a supplementary check for positive log-intensity [35]. Spatial thinning reduced local pseudoreplication [36], and blocked validation respected the spatial structure of the data [37]. Conley HAC was not applied to occurrence because extreme zero inflation made the linear probability formulation unstable. These analyses assessed the stability of directions and rankings; they were not treated as independent significance tests.

2.3.4. Robustness, Geographical Concordance, and Administrative Comparisons

Robustness analyses addressed four sources of uncertainty. Spatial support was varied across 5, 10, and 20 km grids, and GCU dominance thresholds were varied from DOM50 to DOM80 on the 10 km grid. Classification sensitivity was assessed using alternative overlap priorities and binary input specifications. Temporal consistency was examined using the immediate-reversal filter. Initial-stock effects were assessed by dividing turnover and loss by the initial settlement area and gain by the initial non-settlement area for each period. Spatial checks combined Moran’s I, thinning, block bootstrap, restricted permutation, and Conley HAC for positive intensity. We emphasized stable directions and rankings across specifications rather than repeated significance alone.

Independent geomorphology and GLiM data tested specific geographical attributes of each GCU. These were flat land in Lake–Basin, high elevation and relief in Alpine Canyon, carbonate rocks in Karst, compositional centrality in Transition, and mountains in Border–Corridor. Overall association was summarized using Cramér’s V and normalized mutual information (NMI). Class-specific comparisons used differences in proportions, bootstrap confidence intervals, or Jensen–Shannon divergence (JSD). These measures assess convergent geographical concordance and are distinct from profiles of the variables used to construct the classes.

We first quantified GCU mixing within 129 counties and 16 prefecture-level jurisdictions. We then assessed whether GCU membership contributed predictive information beyond administrative categories. Incremental comparisons were conducted separately for each of the three periods using DOM60-assigned 10 km grid cells as prediction units, with period not included as a predictor. Rather than logistic regression or ordinary least-squares regression, we used a nonparametric category-mean prediction framework. For turnover occurrence, the administrative baseline assigned each held-out grid cell the mean occurrence rate of its corresponding county or prefecture estimated from the training data. The enhanced specification added a GCU-specific mean residual correction, also estimated from the training data, to the administrative-category prediction, with resulting probabilities constrained to the interval 0.001–0.999. For positive turnover intensity, only cells with turnover > 0 were retained and annualized turnover intensity was natural-log transformed. The administrative baseline predicted the mean log intensity of the corresponding administrative category in the training data, whereas the enhanced specification added the mean GCU-specific residual correction.

Spatial holdout was implemented using 50 km blocks constructed by aggregating the regular 10 km grid into 5 × 5-cell blocks. These blocks were allocated to five spatial folds with approximately balanced block counts to ensure that observations from the same 50 km block did not occur simultaneously in the training and held-out sets. If an administrative category in a held-out fold was absent from the corresponding training data, its prediction reverted to the training-set global mean; if a GCU category was unseen in the training data, its residual correction was set to zero. Performance was calculated from the pooled out-of-block held-out predictions across the five folds. Turnover occurrence was evaluated using the Brier score, whereas positive turnover intensity was evaluated using mean absolute error on the log scale (log-MAE). Relative improvement was calculated as 100 × (E_admin − E_admin + GCU)/E_admin, where E denotes the corresponding held-out error. County- and prefecture-based baselines were evaluated separately. This comparison was used solely to quantify the out-of-block incremental information contributed by GCU membership beyond administrative categories and was not interpreted as a causal model or as an additional significance test. Complete baseline and enhanced-model errors are reported in Table S8.

The image-checking frame was spatially distributed and stratified by gain, loss, stable settlement, and stable non-settlement. Representative gain and loss locations spanning all three periods were selected from the completed checks. They were compared visually with period-matched Landsat true-color and near-infrared false-color imagery (Figure S2).

3. Results

3.1. GCU Patterns, Reproducibility, and Geographical Concordance

The final classification covered 387,587 valid 1 km cells. Transition occupied 58.00% of the domain, followed by Border–Corridor (15.87%), Karst (14.14%), Alpine Canyon (7.89%), and Lake–Basin (4.10%). Lake–Basin formed discrete patches around major lakes, basins, and broad valleys, whereas Karst was concentrated in eastern and southeastern Yunnan. Alpine Canyon occurred mainly in Diqing, Nujiang, and northwestern Lijiang. Border–Corridor followed the western, southwestern, and southern international boundaries, and Transition covered much of central and inland Yunnan (Figure 3).

Figure 3. Spatial distribution of the five geographical context units and their relationship with administrative boundaries.

Rerunning the four fixed binary input layers with the prescribed priority reproduced all 387,587 cells of GeoType5_V8 without mismatch. This established operational reproducibility; the standardized factor profiles in Figure S1 remained a description of the classification inputs. Independent geomorphology showed moderate overall concordance with GCUs (Cramér’s V = 0.493; NMI = 0.173). The flat-land proportion in Lake–Basin exceeded that in other GCUs by 0.764 (95% confidence interval, 0.718–0.814). The high-elevation and high-relief proportions in Alpine Canyon were higher by 0.655 and 0.660, respectively. GLiM carbonate coverage averaged 66.8% in Karst and 11.3% elsewhere, a difference of 55.5 percentage points and an approximately 5.91-fold ratio. Transition had the most compositionally central geomorphological profile (JSD = 0.119). The mountain component contrast for Border–Corridor was smaller but positive, consistent with a class defined jointly by border location and mountain corridor setting (Figure 4).

Figure 4. External geographical concordance of GCUs assessed using independent geomorphology, GLiM lithology, and compositional centrality. These comparisons measure agreement with geographical reference data rather than classification accuracy against a five-class ground-truth map.

GCUs also captured variation within administrative jurisdictions. Of 129 counties, 68.2% contained at least two GCUs and 22.5% contained at least three. All 16 prefectures contained at least two GCUs, and 93.8% contained at least three. Median dominant-GCU shares were 91.7% for counties and 74.6% for prefectures. In 50 km spatial-block holdouts, adding GCUs to prefecture baselines reduced occurrence Brier scores by 2.06–2.44% and positive-intensity log-MAE by 3.84–6.25%. Corresponding improvements for county baselines were 1.37–1.74% and 2.51–3.81% (Table S8). These consistent but modest gains supported the use of GCUs as a complement to administrative categories.

3.2. Provincial Settlement Stock, Net Change, and Turnover

Settlement area increased from 1355.0823 km² in 1990 to 1402.9839 km² in 2000 and 1759.4595 km² in 2010, then declined to 1739.3742 km² in 2020. The net increase of 384.2919 km² represented 28.36% of the 1990 stock. Across adjacent periods, cumulative gains of 1037.3238 km² and losses of 653.0319 km² yielded 1690.3557 km² of baseline turnover. Conservative immediate-reversal filtering reduced this total to 1117.1556 km². Turnover was therefore 4.40 times the long-term net change under baseline accounting and 2.91 times after filtering (Table 3).

Table 3. Provincial settlement stock and change derived from 30 m data within the area valid in all four years.

Period/Year	Start Stock (km2)	End Stock (km2)	Gain (km2)	Loss (km2)	Net (km2)	Turnover (km2)
1990	—	1355.0823	—	—	—	—
1990–2000	1355.0823	1402.9839	74.4264	26.5248	+47.9016	100.9512
2000–2010	1402.9839	1759.4595	514.8000	158.3244	+356.4756	673.1244
2010–2020	1759.4595	1739.3742	448.0974	468.1827	−20.0853	916.2801
1990–2020 cumulative events	—	—	1037.3238	653.0319	+384.2919 *	1690.3557

Note: The common valid area is 383,617.5291 km². * Long-term stock change. Cumulative gains, losses, and turnover sum events across adjacent periods and may count repeated conversions of the same pixel. Conservative turnover after immediate-reversal filtering was 9.5454, 408.0537, and 699.5565 km² across the three periods, totaling 1117.1556 km².

Baseline turnover rose from 100.9512 km² in 1990–2000 to 673.1244 km² in 2000–2010 and 916.2801 km² in 2010–2020. The conservative estimates retained this temporal ordering, although filtering retained only 9.46% of early-period turnover. The absolute magnitude for 1990–2000 was therefore particularly sensitive to temporal consistency. The final period remained the most active under both accounting schemes. Gains of 448.0974 km² and losses of 468.1827 km² produced a net change of only −20.0853 km², which shows how offsetting transitions obscured substantial turnover. After 2000, high-turnover cells formed broader concentrations in central and eastern Yunnan and at selected border nodes. Alpine Canyon remained dominated by zero or very low turnover (Figure 5).

Figure 5. Mean annual rural settlement turnover intensity on the 10 km grid, 1990–2020. The cumulative panel annualizes the sum of adjacent-period conversion events over 30 years. It does not represent a fourth independent interval or a direct endpoint transition.

3.3. Settlement State, Net Balance, and Turnover Across GCUs

GCU rankings differed among settlement state, net balance, and turnover on the common DOM60 grid (Table 4). Lake–Basin had the highest 2020 settlement density (5.3122%), followed by Karst (0.3819%), Transition (0.3320%), Border–Corridor (0.1798%), and Alpine Canyon (0.0007%). Transition contributed the largest absolute net gain (+228.7458 km²), partly reflecting its extensive area. By contrast, Lake–Basin had the highest median positive ATI in 2010–2020 (0.2312% yr⁻¹). The corresponding values were 0.0176% yr⁻¹ for Karst, 0.0171% yr⁻¹ for Transition, and 0.0111% yr⁻¹ for Border–Corridor. Alpine Canyon had only two positive cells and was not ranked. Absolute contributions and area-normalized intensity thus described different aspects of geographical differentiation.

Table 4. Settlement state, net change, and cumulative turnover across GCUs on the DOM60 grid.

Gcu	DOM60 N	2020 Density (%)	Net 1990–2020 (km2)	Cumulative Gain (km2)	Cumulative Loss (km2)	Turnover Events (km2)
Lake–Basin	51	5.3122	+10.1484	138.2346	128.0862	266.3208
Karst	244	0.3819	+20.9313	46.9296	25.9983	72.9279
Transition	2171	0.3320	+228.7458	461.4210	232.6752	694.0962
Alpine Canyon	230	0.0007	+0.1431	0.1431	0.0000	0.1431
Border–Corridor	586	0.1798	+40.0968	60.8751	20.7783	81.6534

Note: All values refer to the DOM60 grid. Cumulative gains, losses, and turnover sum events across the three adjacent periods. The 2010–2020 median positive ATI values were 0.2312, 0.0176, 0.0171, and 0.0111% yr⁻¹ for Lake–Basin, Karst, Transition, and Border–Corridor, respectively. Alpine Canyon was not ranked because only two cells had positive turnover.

Between-GCU differences in occurrence were similar across periods, with Cramér’s V values of 0.2711, 0.2709, and 0.2675 (Table 5). Block-restricted permutations gave p = 0.0001 in each period, which supports the persistence of differentiation after accounting for local spatial structure. Occurrence rose from 98.0% to 100% in Lake–Basin, 31.1% to 39.8% in Karst, 33.5% to 41.5% in Transition, and 22.2% to 33.3% in Border–Corridor. Alpine Canyon remained at 0%, 0%, and 0.9% across the three periods. Its defining pattern was therefore sparse occurrence rather than a reliably estimated low positive intensity.

Table 5. Primary two-part statistical tests for GCU differentiation in turnover.

Period	Cramér’s V	Occurrence Χ2	50 km Block p	Ε2	Positive N	KW H (df)	KW p
1990–2000	0.2711	241.184	0.0001	0.0966	984	97.705 (3)	<0.001
2000–2010	0.2709	240.913	0.0001	0.0491	1213	62.330 (3)	<0.001
2010–2020	0.2675	234.824	0.0001	0.1061	1245	135.555 (4)	<0.001

Note: Occurrence inference uses 9999 permutations restricted to 50 km spatial blocks. Conditional positive intensity uses tie-corrected Kruskal–Wallis tests and ε² = (H − k + 1)/(N − k), with Dunn–Holm post hoc comparisons. H is the test statistic, k the number of non-empty groups, and N the positive-cell count. Sparse Alpine Canyon observations do not support an intensity ranking.

Differentiation in conditional positive intensity varied more strongly over time (Figure 6). Effect sizes were ε² = 0.0966, 0.0491, and 0.1061, with the weakest separation in 2000–2010 and the strongest in 2010–2020. Kruskal–Wallis tests gave p < 0.001 in all periods. Lake–Basin median positive ATI increased from 0.0205 to 0.1026 and 0.2312% yr⁻¹, remaining above the other major GCUs. Dunn–Holm comparisons most consistently distinguished Lake–Basin from other contexts, whereas Karst, Transition, and Border–Corridor were weakly separated in 2000–2010 (Table S10).

Figure 6. Two-part comparison of rural settlement turnover across GCUs: (a–c) occurrence and (d–f) conditional positive intensity.

3.4. Sources of Settlement Gain and Destinations of Settlement Loss

Cropland was the dominant source of settlement gain in every period, contributing 72.83%, 69.52%, and 74.44% (Table 6). It supplied 71.88% of cumulative gains, followed by grassland (13.03%) and forest (12.01%). Other classes contributed relatively little. Loss destinations changed more markedly over time. The cropland share declined from 79.02% to 55.39% and 47.12%, while other built-up land increased from 2.00% to 13.82% and 37.86%. Across all periods, cropland received 50.42% of losses and other built-up land received 30.58% (Figure 7 and Figure 8). These shifts describe changing transition composition without identifying the policies or management processes responsible.

Table 6. Principal sources of settlement gain and destinations of settlement loss at the provincial level.

Period	Gain from Cropland (%)	Gain from Forest (%)	Gain from Grassland (%)	Loss to Cropland (%)	Loss to Forest (%)	Loss to Grassland (%)	Loss to Other Built-Up (%)
1990–2000	72.83	20.19	6.35	79.02	8.13	8.32	2.00
2000–2010	69.52	11.08	15.90	55.39	15.22	8.26	13.82
2010–2020	74.44	11.73	10.84	47.12	6.20	6.32	37.86
1990–2020 cumulative	71.88	12.01	13.03	50.42	8.47	6.87	30.58

Note: Source percentages use total gain as the denominator, and destination percentages use total loss. The table shows selected categories; water, unused land, and other built-up gain sources are retained in the complete accounting.

Figure 7. Dominant source classes of settlement gain and destination classes of settlement loss on the 10 km grid. “Mixed” indicates that no source or destination class reached a 50% share of the corresponding transitions. This threshold concerns transition composition, separately from GCU assignment. Cumulative panels sum adjacent-period events rather than direct 1990–2020 endpoint transitions.

Figure 8. Source and destination composition of rural settlement transitions at provincial and GCU levels.

Transition composition differed among GCUs. In Lake–Basin, cropland supplied 81.1% of cumulative gains, while losses went mainly to cropland (47.8%) and other built-up land (43.8%). Gains in Karst and Transition were also dominated by cropland (68.5% and 72.1%, respectively), with additional contributions from forest and grassland. Border–Corridor differed: cropland supplied 43.3% of gains and forest 39.6%, while these classes received 44.9% and 31.8% of losses, respectively. Alpine Canyon recorded only 0.1431 km² of gain and no losses, which precludes substantive interpretation of its percentage composition.

3.5. Robustness and Sensitivity

The temporal ranking of mean annual turnover remained 2010–2020 > 2000–2010 > 1990–2000 across the 5, 10, and 20 km grids. Raising the GCU dominance threshold reduced the assigned-cell counts from 3745 at DOM50 to 3282, 2778, and 2277 at DOM60, DOM70, and DOM80, respectively. The principal geographical contrasts retained their direction. Alternative overlap priorities and classification specifications also preserved the core intensity ordering (Table 7).

Table 7. Robustness and sensitivity of turnover estimates and GCU comparisons.

Robustness Dimension	Specification/Metric	1990–2000	2000–2010	2010–2020	Overall Interpretation
Spatial support	Mean annual turnover across 5/10/20 km supports (% yr−1)	0.00252–0.00259	0.01759–0.01806	0.02341–0.02368	Temporal ordering retained at all three supports
GCU dominance threshold	DOM50/DOM60/DOM70/DOM80 assigned N	—	—	—	3745/3282/2778/2277; main geographical contrasts retained
GCU specification	Overlap priority and alternative operational specifications	—	—	—	Core intensity ordering retained
Temporal consistency	Conservative turnover after immediate-reversal filtering (km2; retained %)	9.5454 (9.46%)	408.0537 (60.62%)	699.5565 (76.35%)	Ranking retained; early-period absolute magnitude is most sensitive
Temporal consistency	Conservative Cramér’s V/positive-intensity ε2	0.281/0.100	0.267/0.027	0.267/0.112	Occurrence differentiation retained; positive-intensity separation weakest in 2000–2010
Initial-stock normalization	Turnover/initial settlement stock (% yr−1); positive-intensity ε2	0.745/0.033	4.788/0.011	5.201/0.035	Temporal increase and GCU differentiation persist after stock normalization
Spatial dependence	Queen Moran’s I: occurrence/positive intensity	0.386/0.221	0.403/0.304	0.394/0.313	Spatial dependence requires explicit treatment
20 km spatial thinning	Cramér’s V/ε2; significant-replicate frequency	0.301/0.089; 100%/99.5%	0.301/0.047; 100%/92.3%	0.297/0.125; 100%/100%	Primary thinning stability check
30 km spatial thinning	Cramér’s V/ε2; positive-intensity significant frequency	0.310/0.076; 64.8%	0.311/0.044; 50.0%	0.308/0.117; 94.8%	Stress test; small retained samples
50 km block bootstrap	Cramér’s V/ε2	0.274/0.097	0.274/0.053	0.270/0.108	Effect sizes close to primary estimates
Blocked occurrence permutation	50 km blocks; 9999 permutations	χ2 = 241.184; p = 0.0001	χ2 = 240.913; p = 0.0001	χ2 = 234.824; p = 0.0001	Occurrence differentiation retained
Conley HAC	Positive log-intensity; 50/100/150 km cutoffs	Supported	Supported	Supported	Supplementary intensity check; not applied to occurrence
DOM60 mixed-grid audit	Occurrence: assigned/mixed (%)	29.98/39.86	36.96/47.38	37.93/48.79	Mixed cells are more active; excluded from GCU comparisons
DOM60 mixed-grid audit	Mean ATI: assigned/mixed (% yr−1)	0.00201/0.00510	0.01529/0.03092	0.01861/0.04643	DOM60 underrepresents activity at mixed interfaces
Overall assessment	Stable vs. sensitive findings	Lowest turnover	Weakest positive-intensity separation	Highest turnover; strongest positive-intensity separation	Lake–Basin remains active and Alpine Canyon sparse; early-period magnitude is sensitive

Note: Spatial-support comparisons use KEEP20 cells. DOM denotes the minimum GCU area share required for assignment. The DOM60 audit compares 3282 assigned and 705 mixed cells. Immediate-reversal filtering tests temporal sensitivity without classifying every reversal as error. Thinning frequencies describe stability across dependent subsamples; 30 km thinning is a small-sample stress test. Cumulative turnover sums conversion events and does not measure unique changed area.

DOM60 excluded 705 of 3987 KEEP20 cells (17.7%) from between-GCU inference. These mixed cells were more active than the assigned cells in every period. Their occurrence proportions ranged from 39.86% to 48.79%, compared with 29.98% to 37.93% among assigned cells; the mean ATI was also consistently higher (Table 7). The DOM60 comparisons therefore represented areas clearly dominated by one context and underrepresented activity at geographical interfaces. Mixed cells remained in provincial maps but were not forced into a single GCU.

Temporal filtering changed absolute turnover more than its temporal ranking. Retained turnover increased from 9.46% in 1990–2000 to 60.62% in 2000–2010 and 76.35% in 2010–2020. Stock-normalized annual turnover also increased across periods, from 0.745% to 4.788% and 5.201%. The corresponding positive-intensity effect sizes were ε² = 0.0331, 0.0109, and 0.0351, respectively. Geographical differentiation therefore persisted after accounting for initial settlement stock, with the weakest separation again in 2000–2010.

Queen-contiguity Moran’s I ranged from 0.386 to 0.403 for occurrence and from 0.221 to 0.313 for positive intensity, indicating substantial spatial dependence. Under 20 km thinning, occurrence differences remained significant in every replicate; positive-intensity differences remained significant in 99.5%, 92.3%, and 100% of replicates. The 50 km block bootstrap produced effect sizes close to the primary estimates. At 30 km, some intensity comparisons weakened as retained samples became smaller. This more restrictive thinning was therefore treated as a stress test. Across the tested specifications, Lake–Basin retained high occurrence and intensity, Alpine Canyon remained sparse, and positive-intensity separation was weakest in 2000–2010.

4. Discussion

4.1. Information Revealed by Separating Settlement State, Balance, Process, and Interface

The framework distinguishes four aspects of rural settlement change that answer different questions. Density describes settlement concentration, net change records the balance of gains and losses, and turnover measures their combined magnitude. Source–destination composition identifies the land use categories connected by these transitions. This separation accords with global evidence of gross change exceeding net change [14,15] and with intensity and quantity–exchange approaches to land-change analysis [30,31].

The distinction was clearest in 2010–2020, when a near-zero net change coexisted with the highest turnover under both baseline and conservative accounting. Filtering reduced estimated event areas substantially but retained the temporal ranking and the excess of cumulative turnover over the long-term net change. These relationships were more stable than the absolute turnover estimates, particularly for the earliest period. The main inference is therefore about the information concealed by offsetting transitions, subject to uncertainty in their mapped magnitude.

This distinction is relevant to mountain regions where scarce land, livelihood change, accessibility, and environmental constraints can produce local gains and losses simultaneously [1,8,9]. Comparing stock, net balance, and turnover on a common spatial support makes those contrasting trajectories visible. It also prevents a single label, such as growth or shrinkage, from standing in for the full pattern of land conversion. Identifying the socioeconomic mechanisms behind that pattern requires additional evidence.

4.2. Geographical Context as a Complement to Administrative Comparisons

GCUs provide a geographical layer that can complement the jurisdictions responsible for planning and public services. Scale-mismatch research shows why aggregation matters when environmental processes and governance boundaries do not align [12]. In Yunnan, multiple GCUs occurred in every prefecture and in most counties. Administrative averages could therefore combine distinct topographic, hydrological, karst, and border settings. Retaining both representations allows regional diagnosis to reflect these differences while remaining connected to the institutions that implement decisions.

Three distinct checks supported this complementary role. Exact reproduction from archived inputs established that the classification could be rerun. Independent geomorphology and lithology supported its geographical interpretation, although overall concordance was moderate and topographic agreement for Border–Corridor was partial. Spatial-block holdouts then showed modest, consistent predictive gains from adding GCUs to the administrative categories. Reproducibility, geographical concordance, and additional predictive information address different questions; none alone establishes the full validity of the classification.

The GCU profiles also resist interpretation as a single development gradient. Lake–Basin combined high density and turnover intensity but did not contribute the largest absolute net gain. Transition’s larger totals partly reflected its extensive area, whereas Alpine Canyon was characterized by sparse occurrence and insufficient positive observations for intensity ranking. Moreover, mixed cells were more active than assigned cells. Comparisons among dominant contexts therefore need to be read alongside evidence from their interfaces, which DOM60 deliberately excludes from between-GCU inference.

4.3. Land Use Transitions and Differentiated Planning Assessment

Source–destination analysis clarifies what similar turnover magnitudes can mean for different land systems. Cropland remained the main source of settlement gains, but other built-up land became a greater destination for settlement losses. This latter transition became increasingly prominent even as the cropland share of losses declined. National research has documented the relationship between rural residential expansion and cultivated-land occupation [5]. Our results indicate that monitoring gains and losses together reveals relationships that expansion-focused assessments can miss. Land-class transitions alone cannot determine whether these changes resulted from reclamation, urban absorption, or particular policies.

GCU differences provide a spatial basis for follow-up assessment. Lake–Basin warrants attention to existing settlement land, reuse opportunities, and cropland–built-up transitions. Karst calls for assessment of settlement suitability, soil and water constraints, and transitions involving forest and grassland. Transition requires monitoring across its extensive, low-density settlement landscape. Sparse Alpine Canyon settlements warrant locally grounded hazard and environmental assessment, while Border–Corridor calls for monitoring of corridor nodes and forest–settlement transitions. These are screening priorities derived from observed patterns, rather than parcel-level prescriptions (Table 8).

Table 8. GCU-specific priorities for follow-up planning assessment based on the observed settlement-change profiles.

GCU	Observed Profile	Principal Land Transitions	Follow-Up Priority	Interpretive Boundary
Lake–Basin	Highest 2020 density; turnover occurrence ≈100%; highest positive intensity	Gains mainly from cropland; losses mainly to cropland and other built-up land	Assess existing settlement land, reuse opportunities, and cropland–built-up transitions	No attribution to urban absorption or renewal policy
Karst	Low density; moderate and rising occurrence; positive intensity below Lake–Basin	Cropland dominant; forest and grassland also involved	Assess settlement suitability and transitions involving forest and grassland	Does not measure ecological condition or restoration effect
Transition	Largest absolute 1990–2020 net gain; extensive low-density matrix	Cropland dominant; forest/grassland secondary	Monitor dispersed growth and coordinate services with land use	Absolute gain partly reflects large territorial extent
Alpine Canyon	Near-zero density and structural sparsity; 0/0/2 positive cells	Too few events for compositional interpretation	Assess local hazards and environmental constraints	Do not rank conditional turnover intensity
Border–Corridor	Low density; occurrence rises over time; moderate cumulative turnover	Forest involvement substantially higher than in Lake–Basin	Monitor corridor nodes and forest–settlement transitions	No causal estimate of corridor or border policy effects

Note: The matrix supports regional screening and follow-up assessment. It does not estimate policy effects, household behavior, land rights, or parcel-level planning outcomes.

Existing mechanism studies provide hypotheses for investigating these patterns. Research in rural China has examined the environmental effects of residential expansion [38], land-policy and construction-land transitions [39,40,41,42,43], and social connections, livability, and spatial governance [44,45,46]. Together, these studies indicate that population, institutions, and social organization can shape settlement adjustment. Testing such explanations here would require household surveys, land-rights information, approval records, or a suitable causal design. Those data were outside the present analysis, so transition proportions cannot identify specific policy effects.

4.4. Robustness, Limitations, and Transferability

This analysis separates the spatial roles of change detection, context construction, and statistical comparison. Its principal temporal ordering and geographical contrasts persisted across the tested grids, dominance thresholds, classification specifications, and spatial-dependence checks. Initial-stock normalization also retained geographical differentiation. However, the higher activity among the 705 mixed cells limits the extent to which DOM60 comparisons represent the province’s geographical interfaces. This exclusion is a defined feature of the comparison, not a random loss of observations.

The measurements impose further limits on interpretation. Cumulative turnover counts events across intervals and may count repeated conversion of the same pixel. The 30 m maps resolve regional land use transitions, not individual buildings, parcels, or changes within villages. Source–destination composition identifies land use connections without establishing their social, policy, or ecological causes. Sparse positive observations in Alpine Canyon likewise support occurrence-based interpretation but not a reliable intensity ranking.

Temporal consistency remains a substantial limitation on absolute event areas. The immediate-reversal filter retained only 9.46% of the baseline turnover in 1990–2000, even though temporal ordering remained unchanged. The filtered values are sensitivity estimates, not independently validated corrections, because genuine reversals may also be excluded. We therefore report both estimates and place greater weight on patterns that persist across them. Landsat checks provide independent visual evidence for selected locations but do not estimate map-wide accuracy. Area-adjusted estimates would require a probability reference sample and an error-matrix design [47].

Application elsewhere requires rebuilding the analytical framework around local geography. Context definitions, thresholds, and assignment priorities should reflect the processes and data of the target region. Stock, net balance, turnover, and transition composition can then be compared on consistent spatial supports and tested for temporal and spatial sensitivity. A European review of abandoned agricultural land similarly emphasizes the joint influence of biophysical conditions, institutions, and socioeconomic processes on subsequent trajectories [48]. The potentially transferable contribution is this analytical sequence; the five Yunnan classes require independent reassessment before use in other mountain regions.

5. Conclusions

This study demonstrates that combining cross-administrative geographical context units (GCUs) with settlement state, net balance, turnover, and source–destination composition provides a more complete regional diagnosis of rural settlement change than administrative aggregation or net change alone. In Yunnan, rural settlement area increased by 384.29 km² between 1990 and 2020, whereas cumulative turnover reached 1690.36 km² under baseline accounting and 1117.16 km² under the conservative temporal-consistency filter. Both estimates retained the temporal ordering of 2010–2020 > 2000–2010 > 1990–2000, showing that substantial bidirectional land conversion can coexist with limited net change.

Geographical differentiation was strongly metric-dependent. Lake–Basin exhibited the highest settlement density and turnover intensity, Hill–Mountain Transition contributed the largest absolute net gain, and Alpine Canyon was characterized by structural sparsity. Cropland was the dominant source of settlement gains, whereas other built-up land became an increasingly important destination for settlement losses. Together, these contrasts show that settlement stock, net balance, conversion activity, and land use interfaces capture different dimensions of rural settlement dynamics.

The main contribution of the framework lies in integrating locally defined geographical contexts, bidirectional change accounting, and explicit robustness assessment within a common spatial analytical structure. The GCUs are intended to complement rather than replace administrative units in regional monitoring and planning assessment. Application beyond Yunnan requires locally appropriate context definitions and independent evaluation, and the framework should be interpreted as a diagnostic tool rather than as evidence of policy causality or ecosystem condition.