Land-Use Evolution and Multi-Scenario Spatial Divergence in Bays: A Case Study of Haitan Bay and Tannan Bay, Pingtan Island, China

Abstract

Conflicts between development and conservation on tourism-driven islands are concentrated in bay-adjacent terrestrial areas, yet land-use evolution and scenario responses are rarely compared among functionally distinct bays on the same island. This study examined Haitan Bay and Tannan Bay on Pingtan Island using 2 m remote-sensing images from 2008, 2016, and 2024, landscape metrics, and a coupled System Dynamics–Patch-generating Land Use Simulation (SD–PLUS) model. Natural Development (ND), Ecological Restoration (ER), and Economic Development (ED) scenarios were simulated for 2035; ER and ED outputs were further compared pixel by pixel with Minimum Mapping Unit (MMU) sensitivity testing. From 2008 to 2024, Construction Land expanded and became more aggregated in both bays, mainly at the expense of Cropland and Grassland. Relative to 2024, Construction Land increased by 1.74% and 3.97% under ER in Haitan Bay and Tannan Bay, versus 12.27% and 10.89% under ED. New Construction Land mainly extended around existing built-up areas in Haitan Bay but clustered along transportation corridors in Tannan Bay. Scenario-consistent Areas accounted for 87.56% and 89.29%, while some patch metrics were MMU-sensitive. These results reveal distinct land-use responses between Haitan Bay and Tannan Bay and support differentiated spatial planning for the two bay-adjacent terrestrial areas.

1. Introduction

Compared with continental regions, islands have limited land resources and relatively sensitive ecosystems. Tourism development, infrastructure construction, and urban expansion are often concentrated within narrow coastal zones, intensifying conflicts between conservation and development [1,2]. As human activity intensifies, the limited ecological carrying space of islands faces increasing development pressure. Urban expansion commonly occupies farmland, Forest Land, and coastal wetlands while increasing land-use intensity and landscape fragmentation [3]. Growth in tourism further increases anthropogenic disturbance and ecosystem vulnerability [4,5], thereby promoting continued reorganization of island land-use patterns. Ecological buffers, coastal recreational resources, transportation facilities, and residential development space substantially overlap in coastal zones [6]. Island land-use studies therefore need to identify changes in land quantity, local expansion pathways, and allocation differences among policy scenarios. Such analyses can clarify how development and conservation pressures operate within limited coastal space.

Bays and their adjacent terrestrial areas are important spatial units for examining land-use reorganization and associated landscape responses. Natural ecological spaces, including beaches and coastal shelterbelts, often overlap with transportation corridors, tourism facilities, and urban development in bay-adjacent terrestrial areas. These areas are therefore zones where interactions between natural ecological processes and human development are concentrated [7,8]. However, island-wide changes cannot fully represent local evolution in the terrestrial areas adjacent to individual bays. Comparing multiple bays within the same island context can reveal commonalities and differences in land-use evolution and scenario responses. It can also identify local spatial divergence obscured by island-wide trends. Determining whether functionally distinct bays exhibit distinct historical trajectories, future spatial allocations, and scenario responses requires finer-scale comparison [9].

A range of spatially explicit approaches have been developed for land-use change simulation, providing a relatively mature technical basis for characterizing changes in land demand and associated spatial allocation. CLUE-S simulates regional land-use change and spatial allocation by integrating socioeconomic and biophysical factors [10], while cellular automata models such as SLEUTH focus on urban expansion, and FLUS and PLUS have been further applied to multi-type land-use and multi-scenario spatial simulation [11,12]. These approaches have been applied across different regions and spatial scales. Verburg and Veldkamp (2004) conducted land-use change simulations at the national scale in the Philippines and at the island scale on Sibuyan Island, comparing land-use transitions across spatial scales [13]; Xian et al. (2005) used SLEUTH to simulate urban development in Tampa Bay, USA [14]; Dorning et al. (2015) used the patch-based land-change model FUTURES to compare future spatial patterns under different urbanization scenarios in the North Carolina Piedmont, USA [15]. However, existing studies have largely focused on overall land-use change and future spatial allocation at the scale of islands, cities, or other regional units, with relatively limited attention to differences in land-use evolution and scenario responses among local units within a region under a shared regional development context. Meanwhile, multi-scenario simulation results may also be affected by spatial scale and Minimum Mapping Unit (MMU) settings [16,17,18]. High-resolution data can preserve more local spatial information and small patches, but the retention, merging, or elimination of small patches may vary with the MMU, thereby affecting landscape patterns and associated patch metrics [19]. Therefore, land-demand simulation, spatial allocation, comparison among local units within a region, scenario-divergence identification, and MMU sensitivity testing should be integrated within a unified analytical framework to distinguish relatively stable scenario-allocation differences from local patch characteristics that are sensitive to spatial scale.

Pingtan Island is a representative area where tourism development, infrastructure construction, and ecological conservation are in substantial conflict [20]. Since the Pingtan Comprehensive Experimental Zone was established in 2009, infrastructure construction, tourism development, and spatial-governance policies have jointly reorganized land use across the island [21,22]. Limited coastal space, pressure from tourism expansion, shrinking ecological buffers, and functional differences among bays are also relevant to many tourism-driven islands and coastal regions. Although previous studies have examined land-use change at island or coastal-city scales, functionally distinct bays within the same island have rarely been compared systematically. This study therefore examined the terrestrial areas of Haitan Bay and Tannan Bay on Pingtan Island. We used 2 m-resolution remote-sensing images from 2008, 2016, and 2024 to analyze historical changes in land use and landscape patterns. A System Dynamics–Patch-generating Land Use Simulation (SD–PLUS) coupling framework was then used to simulate land quantity, spatial expansion pathways, and landscape responses from 2025 to 2035. The study addressed two questions. (1) What patterns of land-use evolution and scenario response do different bays exhibit under the same regional development context? (2) How robust across scales are the spatial-allocation differences identified by multi-scenario simulation under different MMU settings? On this basis, differentiated governance implications were proposed by considering the functional roles and local spatial characteristics of the two bays.

2. Materials and Methods

2.1. Study Area

Pingtan (25°15′–25°45′ N, 119°32′–120°10′ E) lies off the central coast of Fujian Province on the western side of the Taiwan Strait. Haitan Island, the main island of Pingtan, covers 267.13 km² and is the fifth-largest island in China. The island has generally low relief and is dominated by marine depositional plains [23]. It has a warm, humid subtropical maritime monsoon climate. Vegetation is dominated by mixed stands of Casuarina equisetifolia and Acacia confusa, and its diverse dunes provide favorable conditions for coastal recreation.

This study selected the bay-adjacent terrestrial areas of Haitan Bay and Tannan Bay as comparative study units (Figure 1). Haitan Bay and Tannan Bay have relatively distinct functional roles in local spatial planning. The Master Plan of Haitan Scenic and Historic Area (2017–2030): Planning Explanatory Report, in coordination with the Master Plan of Pingtan Comprehensive Experimental Zone (2011–2030), designates the Tannan Bay and Haitan Bay clusters within the coastal tourism and recreation space of the main island [24]. The Haitan Bay study area covers 8.44 km² and was a core area of early development on Pingtan Island. Its built-up areas are relatively concentrated, and it has a long history of human use. Its landscape pattern is characterized by intensive anthropogenic disturbance and shoreline artificialization. The Tannan Bay study area covers 10.98 km² and was developed later. It features high-quality beaches, relatively well-preserved coastal shelterbelts, and a favorable ecological background. It therefore provides a representative setting for studying landscape evolution under conflicts between conservation and development [25].

2.2. Data Sources and Processing

To ensure spatial comparability between the two bays and across periods, this study applied a consistent rule to delineate the bay-adjacent terrestrial areas. On the landward side, the island ring road and connected major roads formed a closed boundary, while the 2019 baseline coastline was used uniformly as the seaward boundary. The Master Plan of Haitan Scenic and Historic Area (2017–2030): Planning Explanatory Report uses roads, identifiable landmarks, and physical features as boundary references in spatial delineation and, in some locations, uses the island ring road to define scenic-area boundaries [26]. The Territorial Spatial Ecological Restoration Plan of Pingtan Comprehensive Experimental Zone (2021–2035) further identifies the island ring road as the backbone of the “one-ring” transport–ecological corridor linking major coastal scenic areas and distinctive villages and towns, including Haitan Bay and Tannan Bay [27]. Based on these local spatial-organization characteristics, the island ring road and adjacent major roads provide a continuous, clear, and reproducible landward spatial reference for the two study units. The 2008, 2016, and 2024 land-use data and the 2035 scenario simulations used this fixed study extent, while the 2019 baseline coastline was used to standardize the spatial boundary across periods. Socioeconomic parameters were derived mainly from the Pingtan Statistical Yearbook 2024 and the 2024 Statistical Bulletin on National Economic and Social Development of the Pingtan Comprehensive Experimental Zone. Additional sources were the Territorial Spatial Master Plan of the Pingtan Comprehensive Experimental Zone (2018–2035) and its Report on Sustainable Ecological and Environmental Development (2018–2020). The 2025 Statistical Bulletin on National Economic and Social Development of Fujian Province was also used. The 2025 statistics were used mainly to parameterize future scenarios. Other data sources are listed in

Table 1. Key data and sources.

Category	Data Item	Data Source
Land use data	2008, 2016, and 2024 Remote Sensing Images (2 m)	Island Research Center, Ministry of Natural Resources (MNR)
	2019 Coastline Data	Island Research Center, Ministry of Natural Resources (MNR)
	Ecological Conservation Redline Data	Island Research Center, Ministry of Natural Resources (MNR)
Distance accessibility data	Road Network	OpenStreetMap (https://extract.bbbike.org/ accessed on 27 August 2026)
	Distance to Major Roads	Calculated based on OpenStreetMap
	Distance to Economic Centers	Calculated based on OpenStreetMap
Natural environment data	Digital Elevation Model (DEM)	Geospatial Data Cloud (http://www.gscloud.cn accessed on 27 August 2026)
	Slope	Derived from DEM
	Aspect	Derived from DEM
Socio-economic data	Population Density (100 m)	WorldPop (https://hub.worldpop.org/ accessed on 27 August 2026)

.

Land-use classification was based on three 2 m resolution imagery datasets, all provided by the Island Research Center, Ministry of Natural Resources. The 2008 dataset comprised aerial orthophotographs acquired from 6 to 15 March 2008; the 2016 dataset comprised Gaofen-1 (GF-1) PMS imagery acquired on 19 September 2016; and the 2024 dataset comprised Gaofen-6 (GF-6) PMS orthophotographs acquired on 19 July 2024. Classification accuracy assessment was conducted using same-year high-resolution reference imagery that had not been used in the land-use classification process. The reference imagery for 2016 and 2024 consisted of Google Earth Pro historical images dated 27 March 2016 and 18 December 2024, respectively; for 2008, a 2 m historical remote-sensing image from May 2008 provided by the Island Research Center, Ministry of Natural Resources, was used. The three imagery datasets used for land-use classification were projected to WGS 1984 UTM Zone 50N and geometrically registered in ArcGIS 10.8, yielding a root mean square error of 0.97 m. The images were then clipped to the common study boundaries for land-use classification and landscape pattern analysis.

Using a support vector machine (SVM) in ArcGIS, land use was classified into Forest Land, Construction Land, Cropland and Grassland, Water, Beach, and Bare Land. The rationale for treating cropland and grassland as a single class is described below. Comparison with the land-use category maps from Pingtan’s Third National Land Survey showed that grassland patches in the study area are mostly small and spatially dispersed and are interspersed with cropland patches. For such fragmented and interspersed low-growing vegetation surfaces, crops and grassland can exhibit similar surface-cover characteristics and spectral responses at particular phenological stages, with limited spectral-band differences. Their remote-sensing identification is therefore susceptible to vegetation growth status and image acquisition timing, which limits fine-scale separation based on single-date imagery [28,29]. Meanwhile, the IGBP global land-cover classification includes a Cropland/Natural Vegetation Mosaic class to represent heterogeneous surface units formed by interspersed cropland and natural vegetation, indicating that composite or mosaic classes are an existing mapping approach in remote-sensing land-cover mapping where spatial mixing is pronounced [30]. Given differences in acquisition season and sensor among the three image dates, cropland and grassland were combined into the single class Cropland and Grassland to reduce cross-period confusion between these fragmented and interspersed classes and to maintain class consistency in land-use change analysis and subsequent SD–PLUS simulation. This class was used to characterize the overall area, landscape pattern, and land-use transitions of cultivated surfaces and herbaceous vegetation cover in the study area.

Historical comparisons and scenario simulations in this study focused primarily on changes in area, spatial patterns, and interconversion among Forest Land, Cropland and Grassland, and Construction Land. Although Water, Beach, and Bare Land have distinct surface-cover characteristics, none were treated as an independent land-demand or major scenario-response class within the analytical framework. To maintain a consistent class system across landscape-pattern analysis, classification-accuracy assessment, and SD–PLUS simulation, the three classes were combined as Other Land. This aggregation was applied only in subsequent analyses; the original classification information was retained to identify the historical land-use transitions of Water, Beach, and Bare Land separately. Throughout this study, these terms are capitalized when they refer to the formally defined land-use classes. Cropland and Grassland is treated as a single land-use class in the classification and subsequent analyses.

For classification accuracy assessment, 200 validation points were generated for each bay in each year using stratified random sampling, with the number of points allocated among land-use classes in proportion to their relative areas. The reference land-use class at each validation point was visually interpreted using the reference imagery described above. After quality control, 168 and 179 valid validation points were retained for Tannan Bay and Haitan Bay, respectively, in 2008, while all 200 points were retained for each bay in 2016 and 2024. Overall accuracy (OA) and Kappa coefficients are summarized in

Table 2. Accuracy assessment of remote-sensing classification for Haitan Bay and Tannan Bay in 2008, 2016, and 2024.

Area	Year	Valid Sample Size (n)	OA (%)	Kappa
Tannan Bay	2008	168	92.26	0.89
	2016	200	93.00	0.90
	2024	200	95.00	0.93
Haitan Bay	2008	179	96.09	0.95
	2016	200	94.00	0.92
	2024	200	97.00	0.96

, while class-specific producer’s accuracy (PA), user’s accuracy (UA), and complete confusion matrices are provided in Table S7.

2.3. Methods

The analysis comprised three interlinked analytical levels. (1) Common study boundaries, land-use classes, and landscape metrics were used to compare major land-use change patterns and existing landscape characteristics in both bays from 2008 to 2024. This comparison established a consistent historical baseline for scenario analysis. (2) The SD model projected land demand under ND, ER, and ED. The PLUS model then allocated land use spatially in 2035 using driving factors, neighborhood weights, transition rules, and ecological constraints. (3) ER and ED in 2035 were overlaid pixel by pixel to identify areas where policy-oriented land-allocation differences were most concentrated. Different MMU settings were then used to test the scale robustness of the extent and patch structure of divergent areas. The overall analytical framework is shown in Figure 2.

2.3.1. Historical Land-Use Evolution and Landscape Pattern Analysis

To characterize the spatial distribution of major land-use changes, representative land-use transitions were summarized into four thematic transition groups according to their initial and final land-use classes: Construction Land Expansion, Construction Land Reduction and Vegetation Gain, Other Non-construction Land Transitions, and Forest Reduction and Bare Land Increase. Pixels with the same initial and final land-use class were classified as Unchanged Area. These thematic groups were used primarily to simplify the spatial representation of major land-use transitions and to facilitate comparison of the major land-use change processes between the two bays. Arrows indicate directed transitions from the initial land-use class to the final land-use class, and reverse transitions were treated as distinct land-use change processes. The land-use combinations corresponding to each category are listed in

Table 3. Land-use transition combinations.

Transition Category	Land-Use Transition Combination
Construction Land Expansion	Forest Land → Construction Land; Cropland and Grassland → Construction Land; Water → Construction Land; Beach → Construction Land; Bare Land → Construction Land
Construction Land Reduction and Vegetation Gain	Cropland and Grassland → Forest Land; Construction Land → Forest Land, Cropland and Grassland, Water, Beach, or Bare Land; Beach → Forest Land or Cropland and Grassland; Bare Land → Forest Land
Other Non-construction Land Transitions	Forest Land → Water; Cropland and Grassland → Water; Beach → Water; Bare Land → Cropland and Grassland; Bare Land → Water
Forest Reduction and Bare Land Increase	Forest Land → Cropland and Grassland; Forest Land → Bare Land; Cropland and Grassland → Bare Land; Water → Bare Land; Beach → Bare Land
Unchanged Area	The initial and final land-use classes remained unchanged

Note: Arrows indicate the direction of land-use conversion from the initial land-use type to the final land-use type. Bidirectional conversions between the same pair of land-use types were treated as separate conversion types according to their respective directions.

.

Six class-level landscape metrics—PLAND, PD, LPI, LSI, COHESION, and AI—were calculated using FRAGSTATS 4.2. The land-use rasters for both bays were clipped to their respective study-area boundaries, with pixels outside the study areas set to NoData and excluded from calculation. Patches were identified using the default 8-neighbor rule in FRAGSTATS, whereby pixels of the same land-use class sharing an edge or a corner were treated as belonging to the same contiguous patch, whereas areas separated by other land-use classes or NoData pixels were identified as separate patches. AI was calculated according to the FRAGSTATS definition of like adjacencies, in which pixel adjacency considers shared edges only and excludes diagonal adjacency. These metrics describe the area proportion, number of patches, dominant patch, shape complexity, physical connectivity, and aggregation of each land-use class, respectively; detailed definitions are provided in

Table 4. Landscape pattern indices and their ecological significance.

Landscape Metric	Unit/Range	Meaning and Ecological Interpretation
PLAND	%	Area proportion of a patch class, used to identify the dominant landscape type in the study area.
PD	patches/100 ha	Number of patches per unit area. Under consistent study extent, raster resolution, adjacency rule, and MMU, a higher value generally indicates more patches and provides a relative measure of patch fragmentation.
LPI	%	Area proportion of the largest patch, representing its relative area contribution and dominance in the landscape.
LSI	≥1	Relationship between patch perimeter and area; a higher value indicates a more complex patch shape.
COHESION	0~100	Degree of physical connection among patches; a higher value indicates stronger physical connection.
AI	%	Degree of aggregation among patches of the same class, indicating whether they tend toward clustered or spot-like distributions.

.

2.3.2. SD Model Construction and Validation

System Dynamics (SD) simulates complex social–ecological systems through stock–flow structures and feedback loops. It has been widely applied to regional sustainable development and environmental management [31,32]. An SD model was constructed in Vensim with a temporal extent from 2008 to 2035 and a 1-year time step. Model equations were formulated mainly using lookup and selection functions. Initial stocks were derived from the interpreted 2008 land-use data. Time series for socioeconomic drivers and related historical parameters were constructed and calibrated using statistical yearbooks and bulletins. Interpreted land-use areas for 2016 and 2024 served as historical observations for evaluating the model’s reproduction of past changes in land quantity. The ecological carrying capacity threshold, Maximum Construction Land Capacity, and land-use conversion regulation parameters were determined from planning documents, historical ranges, and previous studies. Parameter sources and calibration procedures are summarized in Table S6. All future scenarios began from the 2024 land-use state and were simulated to 2035. Because the two bays differ substantially in ecological background and functional role, they shared the same core stock–flow framework but used differentiated flow functions and parameters.

Forest Land provides important carbon-storage and ecological-regulation functions, and changes in its pattern are closely related to forest ecosystem resilience [33,34]. Cropland and Grassland is an important open, non-construction land-use class in the study area. Changes in its area and spatial pattern can reflect the overall adjustment of cultivated surfaces and herbaceous vegetation cover during land development and vegetation restoration [35]. Construction Land concentrates population, industry, and infrastructure activities [36]. Bare Land provides relatively limited integrated land-use functions and ecosystem services [37].

Construction Land Area (CLA), Forest Area (FA), and Cropland and Grassland Area (CGA) were defined as the core stocks. Potential demand for construction land expansion was determined jointly by Basic Expansion Demand (BED), Integrated Tourism Development Factor (ITDF), and Maximum Construction Land Capacity (MCLC). Within the SD–PLUS coupling framework, the influence of tourism development is represented mainly through ITDF in the SD land-demand component, which captures its effect on the demand for Construction Land. Conversion of Forest Land to Construction Land was constrained by both the Ecological Space Constraint Factor (ESCF) and Land Use Conversion Regulation Intensity (LUCRI). ESCF represents the constraint imposed by Forest Area relative to the Ecological Carrying Capacity Threshold. Other Land, comprising Water, Beach, and Bare Land, was treated as an area-balancing term and a source of conversion. The model structure is shown in Figure 3, and variable definitions, units, and complete equations are provided in Table S1.

Model validation for the SD simulation comprised a historical fit test and sensitivity analysis. The historical fit test assessed model reliability by comparing simulated results with observed historical data and evaluating the degree of fit. Relative error was calculated using Equation (1):

$$R{E}_{i,t}=\left|{\displaystyle \frac{S_{i,t}-H_{i,t}}{H_{i,t}}}\right|\times 100\%$$

(1)

where S_i,t and H_i,t are the simulated and historically interpreted areas, respectively, for land-use class i in year t. Here, i represents CLA, FA, or CGA, and t represents 2016 or 2024. Areas are expressed in km², and the relative error was calculated separately for each land-use class in Haitan Bay and Tannan Bay.

Sensitivity analysis evaluated model robustness by varying key parameters and examining how changes in these inputs altered the forecasts. The sensitivity coefficient was calculated using Equation (2):

$$S{C}_{X,Y}(t)={\displaystyle \frac{\left|\Delta Y/Y(t)\right|}{\left|\Delta X/X(t)\right|}}$$

(2)

In Equation (2), SC_x,y(t) is the sensitivity coefficient of variable Y with respect to parameter X at time t. ΔX and ΔY denote the changes in parameter X and output variable Y caused by the parameter perturbation, respectively. When SC_x,y(t) < 1, the parameter is considered insensitive, indicating that the model remains relatively robust to variation in that input.

2.3.3. Multi-Scenario Settings

Policy choices inherently involve trade-offs among economic development, ecological conservation, and social benefits [38]. Multi-scenario land-use simulation has been widely used to evaluate the ecological effects of alternative development strategies in coastal cities [39]. This study established the Natural Development Scenario (ND), Economic Development Scenario (ED), and Ecological Restoration Scenario (ER). ND extended historical trends and existing regulation intensity and served as the baseline. ED increased BED and ITDF and raised the loss rates of Forest Land and Cropland and Grassland to represent stronger development demand. ER reduced construction demand and strengthened ecological protection through lower land-loss rates and higher Ecological Restoration Rate (ERR) or Planning-period Restoration Area Cap (PRAC), depending on the bay. Both bays followed the same scenario logic, but parameter values reflected their respective historical ranges, construction capacity, ecological background, and planning constraints. For the post-2024 simulations, the parameter settings in Table S1 correspond to the ND baseline, while scenario-specific adjustments under ED and ER are provided in Table S2.

2.3.4. PLUS Spatial Simulation and Validation

The Patch-generating Land Use Simulation (PLUS) model comprises the Land Expansion Analysis Strategy (LEAS) and CA based on multi-type random patch seeds (CARS). LEAS extracts expansion pixels for each land-use class from changes between two periods. It then uses a random forest to estimate the relative contributions of driving factors and growth probability. CARS generates future land-use patches by combining land demand, growth probability, neighborhood effects, transition rules, and spatial constraints. In the LEAS random-forest training, the number of regression trees was set to 20 and the sampling rate was set to 0.01. CARS simulations used a 3 × 3 neighborhood, with the patch-generation parameter set to 0.2, the expansion coefficient to 0.5, and the percentage of seeds to 0.005. Neighborhood weights, the land-use conversion cost matrix, and scenario-specific land-use demands were specified accordingly.

Seven driving factors were selected according to geographical characteristics and data availability. They were Population Density, Digital Elevation Model (DEM), Slope, Aspect, Distance to Major Roads, Distance to Economic Centers, and Distance to Coastline. All factors were standardized to the same projection, spatial resolution, and study extent. Distance variables were calculated using Euclidean distance. The ecological conservation redline was included as a restricted development zone in the spatial allocation. As a spatial constraint, the ecological conservation redline restricts where land-use conversions can occur in the scenario simulations and operates independently of the land-use conversion cost matrix. Values of 0/1 in the conversion cost matrix indicate only whether conversion between specific land-use classes is permitted under a given scenario, whereas the ecological conservation redline defines the spatial extent within which these conversions may occur. The driving factors and spatial constraints are shown in Figure 4 and Figure 5, respectively.

The period 2008–2016 was used to parameterize the neighborhood weights (

Table 5. Weights of land-use types in Haitan Bay and Tannan Bay.

Land Use Type	Haitan Bay	Tannan Bay
Forest Land	0.46	0.34
Cropland and Grassland	0.05	0.02
Construction Land	0.44	0.62
Other Land	0.05	0.02

), followed by temporal validation for 2016–2024. Using the actual 2016 land-use pattern as the initial state, the neighborhood weights derived from 2008–2016 were applied to simulate the 2024 spatial pattern, which was then compared with the actual 2024 land-use map. During this temporal validation, the parameterized PLUS settings were not recalibrated. This process provided historical validation support for the subsequent 2035 scenario simulations.

For the 2035 spatial simulations, scenario-specific land demands generated by the SD model were used as the quantity constraints in PLUS. Spatial allocation was further controlled by historical neighborhood effects, land-use transition rules, and spatial constraints, including the ecological conservation redline. Thus, the simulated 2035 land-use patterns reflected the combined effects of scenario-specific land demand and spatial allocation rules. The land-use conversion cost matrix was adapted from previous research [40] to reflect the conservation and development constraints of the study area (

Table 6. Land-use conversion cost matrix under different scenarios.

Scenarios	Natural Development				Ecological Restoration				Economic Development
Land Use Type	a	b	c	d	a	b	c	d	a	b	c	d
a	1	1	1	1	1	0	0	0	1	0	1	1
b	1	1	1	1	1	1	0	0	0	1	1	1
c	1	1	1	1	1	1	1	1	0	1	1	1
d	0	0	0	1	0	0	0	1	0	0	0	1

Note: a, b, c, and d represent Forest Land, Cropland and Grassland, Other Land, and Construction Land, respectively. A value of 0 indicates that the conversion is prohibited, whereas a value of 1 indicates that the conversion is allowed. Rows represent the source land-use types, and columns represent the target land-use types.

).

Simulation performance was evaluated using the Kappa coefficient, Overall Accuracy (OA), Figure of Merit (FoM), Quantity Disagreement (QD), and Allocation Disagreement (AD).

To characterize the local aggregation of major conversion processes, Boolean raster algebra was used to extract pixels that changed into or out of each class from 2024 to 2035. Target conversion pixels were assigned a value of 1, and all other pixels were assigned a value of 0. A focal mean was then calculated using a circular neighborhood with a 40-pixel radius (80 m). Local land-use conversion intensity was defined as the proportion of target conversion pixels among all valid pixels in the neighborhood. The metric ranged from 0 to 1, with higher values indicating greater local concentration of the corresponding conversion. The same raster resolution, neighborhood extent, and calculation method were used for both bays and all three scenarios.

2.3.5. Multi-Scenario Spatial Divergence and MMU Sensitivity Test

To identify spatial land-allocation differences between policy orientations, ER and ED in 2035 were overlaid pixel by pixel. ER and ED were selected for the spatial-divergence analysis because they represent the contrasting ecological-conservation and development-oriented policy settings, whereas ND served as the baseline scenario for comparing overall land-use responses. The 2035 ER and ED maps used in this pixel-by-pixel comparison each represent a single CARS simulation realization. The overlay was divided into four classes according to the land-use type assigned to each corresponding pixel. Scenario-consistent Areas had the same land-use type in both scenarios. Construction-Expansion Divergence Areas comprised pixels assigned as Construction Land under ED but as Forest Land or Cropland and Grassland under ER. Forest-Restoration Divergence Areas comprised pixels assigned as Forest Land under ER but as Cropland and Grassland or Other Land under ED. Other Scenario-Divergent Areas comprised all remaining pixels with different land-use types between the scenarios. Because this last class included land-use combinations with inconsistent directions, it was not a focus of comparisons of construction expansion and forest restoration.

Contiguous patches were identified using the 8-neighbor rule. Area Proportion, Number of Patches (NP), PD, Mean Patch Area (AREA_MN), and Largest Patch Share (LPS) were calculated under three MMU settings. These were No MMU Filtering, ≥9 pixels (0.0036 ha), and ≥25 pixels (0.0100 ha). The MMU settings were selected by jointly considering the original spatial resolution and the analysis scale. Because the Scenario-Divergent Areas were identified by pixel-by-pixel overlay of 2 m rasters, relatively small MMUs were required to reduce the influence of very small pixel clusters on patch analysis while retaining as much spatial detail as possible from the high-resolution data [41,42]. On this basis, a 25-pixel threshold (100 m²) was additionally adopted as a higher MMU level, which is consistent with the minimum mapping area specified for Construction Land and facility agricultural land in the Third National Land Survey of Fujian Province [43]. LPS was calculated from the LPI and PLAND outputs from FRAGSTATS as follows:

$$LP{S}_i={\displaystyle \frac{LP{I}_i}{PLAN{D}_i}}\times 100\%$$

(3)

where $LP{S}_i$ is the largest patch area in scenario-zone class $i$ as a proportion of that class’s total area (%). $LP{I}_i$ is the Largest Patch Index of scenario-zone class $i$ (%). $PLAN{D}_i$ is the proportion of scenario-zone class $i$ in the total study area (%). Area Retention Rate under different MMU settings was calculated as follows:

$$R{R}_m=\left({\displaystyle \frac{A_m}{A_0}}\right)\times 100\%$$

(4)

where $A_m$ is the retained area under MMU setting $m$, and $A_0$ is the unfiltered area.

3. Results

3.1. Model Accuracy Assessment

3.1.1. SD Model Accuracy and Parameter Sensitivity Test

For both bays in 2016 and 2024, the absolute relative errors between simulated and interpreted values for Forest Land, Cropland and Grassland, and Construction Land were below 10% (

Table 7. Relative errors for SD model validation.

Study Area	Land-Use Type	2016	2024
		Relative Error (%)
Haitan Bay	Forest Land	8.09	7.59
	Cropland and Grassland	8.59	8.32
	Construction Land	7.03	0.16
Tannan Bay	Forest Land	8.74	0.36
	Cropland and Grassland	5.46	0.16
	Construction Land	2.63	0.97

). Relative errors ranged from 0.16% to 8.59% in Haitan Bay and from 0.16% to 8.74% in Tannan Bay.

Local robustness around the baseline parameters was tested using ND as the baseline. Each of the 7 key parameters in the models for the two bays was perturbed individually by ±10%. All sensitivity coefficients were below 1 in both bays (

Table 8. Parameter sensitivity test results for the SD model.

Study Area	Parameter	Maximum Sensitivity Coefficient Under a 10% Parameter Decrease	Maximum Sensitivity Coefficient Under a 10% Parameter Increase
Haitan Bay	Basic Expansion Demand	0.0129	0.0130
Tannan Bay		0.0234	0.0232
Haitan Bay	Land Use Conversion Regulation Intensity	0.1000	0.0999
Tannan Bay		0.0479	0.0479
Haitan Bay	Ecological Restoration Rate	0.1356	0.1360
Tannan Bay		0.0191	0.0150
Haitan Bay	Maximum Construction Land Capacity	0.1805	0.1500
Tannan Bay		0.5867	0.4620
Haitan Bay	Cropland and Grassland Loss Rate	0.0964	0.0956
Tannan Bay		0.0938	0.0930
Haitan Bay	Forest Loss Rate	0.0230	0.0229
Tannan Bay		0.0060	0.0060
Haitan Bay	Integrated Tourism Development Factor	0.0280	0.0280
Tannan Bay		0.0136	0.0135

), indicating good local robustness.

Together, the historical simulation accuracy and parameter sensitivity results indicate that the model reproduced changes in land quantity and remained relatively robust to local parameter perturbations. The historical fit therefore supported subsequent scenario-based land-demand projections.

3.1.2. PLUS Model Accuracy and Contributions of Driving Factors

OA values for the PLUS simulation were 0.79 in Haitan Bay and 0.74 in Tannan Bay. The corresponding Kappa coefficients were 0.69 and 0.65, and the FoM values were 0.46 and 0.35, respectively (

Table 9. Accuracy assessment of the PLUS land-use simulation.

Evaluation Indicator	Haitan Bay	Tannan Bay
Overall Accuracy (OA)	0.79	0.74
Kappa coefficient	0.69	0.65
Figure of Merit (FoM)	0.46	0.35
Quantity Disagreement (QD, %)	0.85	0.00
Allocation Disagreement (AD, %)	20.15	26.00

). Quantity Disagreement (QD) was 0.85% in Haitan Bay and 0.00% in Tannan Bay, while Allocation Disagreement (AD) was 20.15% and 26.00%, respectively. Overall, QD was low in both study areas, whereas AD was comparatively high, indicating that the discrepancies in the simulation results were concentrated mainly in the spatial allocation of land-use classes. AD was higher in Tannan Bay than in Haitan Bay.

The relative contributions of the driving factors differed between the two bays (Figure 6). In Haitan Bay, the dominant driving factors differed among newly developed areas of the three main land-use classes. New Construction Land was mainly associated with Distance to Coastline, Population Density, and DEM. New Cropland and Grassland and Forest Land were mainly associated with Distance to Economic Centers, Population Density, and Distance to Coastline. In Tannan Bay, Population Density made the largest contribution to all three classes, followed by Distance to Economic Centers. The remaining factors made generally smaller and relatively similar contributions.

3.2. Land-Use and Landscape Pattern Evolution from 2008 to 2024

From 2008 to 2016, Construction Land Expansion covered a relatively large area in Haitan Bay, with most new Construction Land concentrated in the southern and northern parts of the bay. Construction Land Reduction and Vegetation Gain was comparatively limited. Construction Land Expansion was also pronounced in northern Tannan Bay, while large Unchanged Areas and areas of Other Non-construction Land Transitions remained in the bay-adjacent terrestrial area (Figure 7).

From 2016 to 2024, the extent of Construction Land Expansion and Forest Reduction and Bare Land Increase decreased in both bays relative to the previous period, whereas Construction Land Reduction and Vegetation Gain and Unchanged Areas increased. In Haitan Bay, Construction Land Expansion decreased while Construction Land Reduction and Vegetation Gain increased. The overall extent of land-use conversion contracted in Tannan Bay, although local Construction Land Expansion remained.

3.2.1. Land-Use and Landscape Pattern Evolution in Haitan Bay

Construction Land in Haitan Bay expanded continuously and became more aggregated from 2008 to 2024 (Figure 8). Its PLAND increased from 13.56% to 36.10%, LPI increased from 1.02% to 22.69%, and PD decreased markedly. PLAND of Forest Land first decreased and then increased (53.36–25.02–31.49%), while COHESION remained above 99% throughout. Cropland and Grassland area first increased and then decreased, while both PD and LSI first decreased and then increased (Figure 9).

3.2.2. Land-Use and Landscape Pattern Evolution in Tannan Bay

Tannan Bay underwent more pronounced land-use reorganization from 2008 to 2024 (Figure 10). PLAND of Construction Land increased from 14.99% to 42.54%, and LPI rose from 0.42% to 24.91%. PLAND, COHESION, and AI of Forest Land increased by 27.94%, 1.44%, and 22.48%, respectively. PLAND of Cropland and Grassland decreased from 44.67% to 14.50%, while PD and LSI decreased by 53.84% and 63.87%, respectively (Figure 11).

Overall, both bays showed increased Construction Land area, expansion of dominant construction patches, lower Construction Land patch density, and continued contraction of Cropland and Grassland. Under the same raster resolution and adjacency rule, COHESION of Forest Land remained high in both bays. This pattern indicates that the physical connectivity of Forest Land was generally high. Differences in conversion extent and the location of new Construction Land provided distinct historical baselines for future scenario simulations.

3.3. SD-Based Land-Demand Projections

Under ND, both bays continued their historical trajectories, and Construction Land increased through 2035. Construction Land reached 322.08 ha in Haitan Bay and 492.25 ha in Tannan Bay, while Cropland and Grassland continued to decrease. After an earlier decline, Forest Land in Haitan Bay recovered to 316.31 ha, and that in Tannan Bay increased to 455.02 ha (Figure 12).

Under ER, Construction Land increased less than under the other two scenarios in both bays. Relative to 2024, Construction Land in 2035 increased by 1.74% in Haitan Bay and 3.97% in Tannan Bay. Forest Land reached 323.09 ha and 477.38 ha, respectively. Cropland and Grassland decreased less than under ND and ED (Figure 13).

Under ED, Construction Land increased by 12.27% in Haitan Bay and 10.89% in Tannan Bay. Cropland and Grassland decreased by 20.85% and 24.97%, respectively. Net changes in Forest Land differed in direction: an increase of 2.07% in Haitan Bay and a decrease of 0.17% in Tannan Bay (Figure 13). Detailed annual land-use areas are provided in Table S3.

3.4. Multi-Scenario Spatial Allocation and Landscape Pattern Responses

3.4.1. Multi-Scenario Spatial Allocation of Land Use

Spatially, ER produced more conversions into Forest Land in both bays, with relatively limited losses of Cropland and Grassland and less new Construction Land. Under ED, conversions into Construction Land expanded in both bays and were accompanied by local losses of Forest Land. In net-area terms, Forest Land increased slightly in Haitan Bay but decreased slightly in Tannan Bay under ED.

In Tannan Bay, areas with relatively high local intensity of conversion to Construction Land were distributed mainly along transportation corridors, including the island ring road. Additional occurrences were scattered across the central area. Conversions between Forest Land and Cropland and Grassland were concentrated locally in the northeastern and southwestern areas (Figure 14). In Haitan Bay, areas with relatively high local intensity of conversion to Construction Land appeared mainly along the western, northwestern, and southern margins. Conversions into Forest Land were concentrated in the southeastern and northern areas. Areas with relatively high local intensity of conversion out of Cropland and Grassland were spatially close to areas converted into Construction Land or Forest Land (Figure 15).

3.4.2. Multi-Scenario Landscape Pattern Responses

In 2035, PLAND, LPI, and AI of Construction Land reached their highest values under ED in both bays, whereas PD was highest under ER. PLAND and LPI of Cropland and Grassland were highest under ER and lowest under ED. LSI of Forest Land was lowest under ED in both bays. COHESION exceeded 95 for every class, with small differences among scenarios. Thus, the physical connectivity of patches in each class was generally high under all scenarios (Figure 16).

Inter-bay differences in scenario responses were most evident in the shape complexity of Construction Land and the density and shape complexity of Cropland and Grassland patches. Differences were also evident in the density and aggregation of Forest Land patches. In Haitan Bay, LSI of Construction Land was similar under ND and ED, at 81.40 and 81.05, respectively, and exceeded the ER value of 78.97. In Tannan Bay, LSI of Construction Land was highest under ER (119.10) and lowest under ED (106.78). Under ER, PD and LSI of Cropland and Grassland in Haitan Bay were the lowest among the three scenarios, at 621.02 and 82.43, respectively. In Tannan Bay, both metrics were highest under ER, at 1307.00 and 116.37. For Forest Land in Tannan Bay, PD was lowest and AI was highest under ER, at 641.43 and 91.52. In Haitan Bay, PD of Forest Land was highest under ED (563.75), whereas AI was lowest under ER (93.43) (Figure 16).

3.5. Scenario-Divergent Areas and MMU Sensitivity in the Two Bays

3.5.1. Multi-Scenario Spatial Divergence in the Two Bays

In 2035, Scenario-consistent Areas dominated the land-use patterns of both bays, accounting for 87.56% of Haitan Bay and 89.29% of Tannan Bay. Scenario-Divergent Areas accounted for 12.44% and 10.71%, respectively. Divergent areas occurred mainly around existing built-up areas, near roads, and where Forest Land intersected with Cropland and Grassland (Figure 17). Construction-Expansion Divergence Areas accounted for 3.69% of Haitan Bay and 3.15% of Tannan Bay. Forest-Restoration Divergence Areas accounted for 0.24% and 2.81%, respectively.

Within Construction-Expansion Divergence Areas, land-allocation differences primarily involved the retention or conversion of Cropland and Grassland. Pixels assigned as Cropland and Grassland under ER but as Construction Land under ED accounted for 93.53% in Haitan Bay and 94.09% in Tannan Bay. Thus, scenario-based differences in construction expansion mainly reflected allocation differences between retaining Cropland and Grassland and expanding Construction Land. Other Scenario-Divergent Areas accounted for 8.51% and 4.75% of the two study areas, respectively.

3.5.2. MMU Sensitivity and Relatively Robust Inter-Bay Differences

For Construction-Expansion Divergence Areas, Haitan Bay had a slightly higher Area Proportion under No MMU Filtering than Tannan Bay. Its Area Retention Rates were also slightly higher under the ≥9-pixel and ≥25-pixel settings (

Table 10. Area extent and retention of scenario-divergent areas under different minimum mapping unit settings in Haitan Bay and Tannan Bay.

Scenario-Divergent Areas	Bay	Area Proportion (%)	Area Retention Rate (%)
		No MMU Filtering	≥9 Pixels	≥25 Pixels
Construction-Expansion Divergence Areas	Haitan Bay	3.69	87.04	76.50
	Tannan Bay	3.15	81.70	67.93
Forest-Restoration Divergence Areas	Haitan Bay	0.24	45.16	23.17
	Tannan Bay	2.81	89.93	81.66

Note: MMU denotes the minimum mapping unit. “No MMU Filtering” refers to the original pixel-level overlay result without applying a minimum patch-area threshold. Contiguous patches were identified using the 8-neighbor rule. The ≥9-pixel and ≥25-pixel MMU settings retained patches with areas of at least 0.0036 ha and 0.0100 ha, respectively. Area Retention Rate was calculated as the area retained under each MMU setting divided by the corresponding area in the result without MMU filtering. Complete patch metrics for the three analysis scales are provided in Table S5.

). Spatially, divergent patches in Haitan Bay mainly adjoined the margins of existing built-up areas. In Tannan Bay, they were distributed at multiple locations along roads and in central and peripheral areas. The inter-bay rankings of NP, PD, and AREA_MN changed with MMU settings. Haitan Bay had a higher LPS than Tannan Bay at all three analysis scales (Table S5).

For Forest-Restoration Divergence Areas, the directions of inter-bay differences in AREA_MN and LPS were consistent across MMU settings. Area Retention Rates under the ≥9-pixel and ≥25-pixel settings were 89.93% and 81.66% in Tannan Bay. These values were markedly higher than the corresponding rates of 45.16% and 23.17% in Haitan Bay (Table 10). NP, PD, AREA_MN, and LPS were also higher in Tannan Bay at all three analysis scales (Table S5). Area Retention Rate in Haitan Bay decreased rapidly as the MMU increased.

4. Discussion

4.1. Commonalities and Differences in Historical Landscape Evolution Between the Two Bays

From 2008 to 2024, both Haitan Bay and Tannan Bay showed overall construction land expansion and contraction of Cropland and Grassland. Changes in the landscape metrics indicate that construction land expansion in both bays was accompanied by larger dominant patches and fewer patches per unit area. The Construction Land landscape therefore shifted from the coexistence of numerous patches toward dominance by larger patches. Land-use transition results showed that, under the classification system used in this study, Cropland and Grassland was the main source class for new Construction Land, indicating that conversion from Cropland and Grassland to Construction Land was one of the major pathways of historical land-use reorganization in both bays. Forest Land in Haitan Bay exhibited a clear stage-dependent pattern. Its decrease during 2008–2016 was consistent with the regional context of rapid development, construction, and land-use reorganization in Pingtan [44]. During 2016–2024, Forest Land partially recovered, consistent with regional changes characterized by the gradual progress of afforestation, vegetation restoration, and ecological protection on Haitan Island [22,45].

The inter-bay comparison further indicates that the same land-conversion direction does not necessarily produce the same spatial reorganization. The form of construction expansion was also associated with the existing development base, transportation organization, and structure of convertible land. Haitan Bay was developed earlier, with relatively concentrated existing Construction Land. New Construction Land expanded mainly along the margins of existing built-up areas, largely extending the established development pattern. Tannan Bay was developed later and retained more non-construction space; coastal recreation and tourism functions, together with transport infrastructure such as the island ring road, formed the spatial context for development. New development showed a stronger spatial association with transportation corridors, with localized expansion around multiple nodes in central and peripheral areas. Previous studies have shown that urban proximity and transportation accessibility are important factors associated with the location and form of built-up land growth. Road networks can also strengthen the association of construction expansion with transportation corridors [46,47].

Overall, although Haitan Bay and Tannan Bay share the same island-scale development context and exhibit similar overall directions of land-use conversion, they followed clearly different pathways of spatial reorganization. These inter-bay differences correspond to differences in existing land-use patterns, development stages, and spatial structures, indicating that island-scale changes may obscure spatial heterogeneity among local units within the region. Bay-scale comparison can therefore complement island-wide land-use assessment and provide a basis for interpreting differences in scenario responses among local units and for developing more targeted spatial-management strategies.

4.2. Multi-Scenario Spatial Divergence Between the Two Bays Under Historical Patterns

Previous studies suggest that ecological conservation constraints can reduce construction land expansion and prioritize ecological land allocation. By contrast, development-oriented policies can increase demand for construction space and intensify conversion pressure on non-construction land [11,48]. This study showed similar policy responses, with allocations among Construction Land, Cropland and Grassland, and Forest Land varying by scenario objective. Scenario responses were also closely associated with historical land-use patterns. Haitan Bay had a relatively continuous existing development pattern, and its scenario differences mainly involved the magnitude of expansion along established development pathways. Tannan Bay had a more interspersed land-use pattern and more varied types and locations of convertible space. Existing landscape patterns embody legacies of historical land-use change. Together with transition rules and spatial constraints, these legacies are associated with the spatial allocation of future land demand [49,50,51]. The same scenario rules can therefore produce differentiated responses among bays.

In terms of land allocation, Construction-Expansion Divergence Areas in both bays mainly involved differences between Construction Land and Cropland and Grassland. Direct allocation differences between Construction Land and Forest Land were relatively limited. Land-allocation differences between policy objectives primarily involved the retention or conversion of Cropland and Grassland. They reflected competition between growing construction demand and the retention of non-construction space. Cropland and Grassland was both an important target of future construction expansion and the main land-use class involved in spatial differences between the policy scenarios. Future spatial management should focus on areas with concentrated conversion out of Cropland and Grassland. Construction Land allocation and non-construction space retention should also be coordinated according to the divergence characteristics of each bay.

The MMU sensitivity test further clarified the scale applicability of the identified scenario divergence. NP, PD, and AREA_MN in Construction-Expansion Divergence Areas changed markedly with MMU, and some inter-bay rankings also changed; Forest-Restoration Divergence Areas in Tannan Bay had a high Area Retention Rate, while the direction of inter-bay differences in AREA_MN and LPS remained consistent across MMU settings. These results show that the scale responses of individual metrics differ. Metrics such as patch number, density, and mean area should be interpreted in relation to the MMU setting, whereas the main land-allocation relationships and some inter-bay differences remained consistent across MMU settings, showing relative stability across scales. Smaller MMUs retain more small patches, and specific thresholds should be selected according to the analytical objective and study scale.

4.3. Implications for Differentiated Governance at the Bay Scale

Results from the two bays indicate that bay governance should not apply uniform controls based only on the total area of Construction Land. It should also consider the location of new development, inter-scenario land-allocation differences, and their scale robustness. Scenario-divergent Areas can be used to screen for potential allocation conflicts but should not be treated directly as construction restriction zones or ecological restoration priority areas. Divergent patches retained under all MMU settings can be prioritized for verification against current land use, territorial spatial planning boundaries, and ecological restoration suitability [52,53]. NP, PD, and AREA_MN, which change markedly with MMU settings, should be interpreted using multi-scale results and specific patch locations. Governance measures should not be derived directly from a single metric [54,55].

Existing Construction Land in Haitan Bay is relatively concentrated, and new development occurs mainly around existing built-up areas. Given its combined functions of coastal tourism, urban–scenic-area interaction, and urban public services, the land-use efficiency of existing Construction Land should be improved, with particular attention to coordination between peripheral new development and territorial spatial planning boundaries. Management requirements for Cropland and Grassland areas with concentrated conversion to Construction Land should be determined according to the specific current land-use type, planning boundaries, and corresponding management attributes, so that the composite classification is not used directly as a uniform basis for land-use control. Coastal recreation and tourism are important functions in Tannan Bay [56]. Road accessibility provides favorable conditions for the location of tourism facilities, while beaches, natural shorelines, and ecological restoration spaces require stronger protection [57]. Existing development nodes and road-adjacent areas have accessibility advantages [58,59]. Therefore, development along transportation corridors and expansion toward ecologically sensitive areas should be carefully controlled to reduce potential landscape fragmentation and ecological risks [60].

This study proposes differentiated governance recommendations for the two bays by integrating the identification of multi-scenario spatial divergence and MMU sensitivity testing with local functions and ecological constraints. The approach distinguishes the locations, scale robustness, and practical governance implications of scenario differences. It may inform islands and coastal regions where development and conservation are in substantial conflict.

4.4. Uncertainties and Scope of Applicability

The delineation of the study boundary defines the spatial scope within which the results should be interpreted. The same road–coastline delineation rule was applied to Haitan Bay and Tannan Bay and kept consistent in the 2008–2024 historical analysis and the 2035 scenario simulations, providing a common spatial basis for comparisons between the two bays and across periods. Fixing the 2019 baseline coastline means that actual shoreline advance or retreat and the associated changes in terrestrial extent were not included; the conclusions therefore apply primarily to the bay-adjacent terrestrial areas within these fixed boundaries. In land-use studies of bays and coastal zones, terrestrial extents are commonly defined according to the research objective [61,62], and alternative boundary definitions change the land area included in the analysis and may affect comparisons of landscape patterns [63], while truncation of contiguous patches by landscape boundaries may also affect some metrics [64]. Accordingly, the comparison between the two bays is restricted to a consistent and fixed study boundary, and the results are interpreted primarily as relative differences at this spatial scale.

Although stratified random sampling was used in this study to improve sample allocation across different classes, class-specific accuracy estimates may still be affected by the number and distribution of validation samples. Model error, the spatial allocation process, and the temporal representativeness of parameters further constrain the precision with which the 2035 scenario results can be interpreted. At high spatial resolution, local discrepancies between simulated pixels and observed change locations, as well as slight misalignment of land-use boundaries, are more apparent in pixel-level comparisons [12,65,66]. Error analysis showed that validation errors in this study were manifested mainly as spatial-allocation disagreement, while differences in land-use quantities were relatively small. The model results are therefore more suitable for identifying major land-conversion directions and overall spatial-allocation relationships; uncertainty remains in the interpretation of specific pixel locations and local boundaries. Spatial allocation in the 2035 scenarios is also affected by the stochastic allocation process, and some pixel-level differences in Scenario-Divergent Areas may vary among individual simulation realizations [67,68]. In addition, the PLUS neighborhood weights were derived from land-use changes during 2008–2016 and, after validation for 2016–2024, were applied to the 2035 scenario simulations. The relationships governing future land expansion may change, so this parameterization remains uncertain. The causal structure and policy-regulation parameters of the SD model, together with the land-use transition rules and spatial constraints of the PLUS model, define the conditions of the 2035 scenario simulations. The scenario results therefore represent relative land-system responses under the specified parameters, transition rules, and spatial constraints and should be interpreted as conditional scenario results. Climate and coastal-hazard processes, including typhoons, extreme rainfall, storm surges, and sea-level change, were not incorporated as disturbances in the scenario simulations. These processes may alter the coastal environment, suitability for land development, and ecological-conservation conditions, thereby affecting spatial land-use responses. Future studies could incorporate extreme-climate and coastal-hazard processes into land-use scenarios to assess their effects on land-use responses and spatial allocation in the two bays.

Spatial resolution and MMU settings can also affect the stability of landscape-pattern metrics. High spatial resolution preserves local spatial differences and small-scale patches, whereas different MMU settings alter the retention, merging, or elimination of small patches and thereby affect landscape metrics [19,69,70]. The MMU sensitivity test showed that metrics closely related to patch number and scale, including NP, PD, and AREA_MN, changed markedly with filtering thresholds, and some inter-bay metric rankings also changed. The main land-allocation relationships and some relative inter-bay relationships remained stable across MMU settings, although Area Retention Rate in Scenario-Divergent Areas still varied by bay and divergence type. Inter-bay comparisons of patch number, density, and mean area therefore remain constrained by spatial scale; the spatial distribution of scenario divergence and the main land-conversion relationships provide a more appropriate basis for cross-scenario comparison.

5. Conclusions

This study used 2 m-resolution remote-sensing imagery from 2008, 2016, and 2024 to construct an SD–PLUS coupling framework. The framework supported comparison of historical land-use evolution in Haitan Bay and Tannan Bay on Pingtan Island. It also simulated changes in land quantity, spatial allocation, and landscape responses under ND, ER, and ED in 2035. The main conclusions are as follows:

(1)

From 2008 to 2024, both bays showed expansion of Construction Land and contraction of Cropland and Grassland. Conversion from Cropland and Grassland to Construction Land was one of the major pathways of historical land-use reorganization in both bays. Despite similar overall directions of change, the two bays still differed clearly in their land-use evolution: Forest Land in Haitan Bay exhibited a stage-dependent pattern of decline followed by recovery, whereas Tannan Bay underwent more pronounced overall land-use reorganization, with the contraction of Cropland and Grassland being particularly prominent.

(2)

Under all three scenarios, Construction Land increased and Cropland and Grassland decreased in both bays, although the magnitudes of change differed. Under ER, Construction Land showed the smallest increase, Cropland and Grassland showed the smallest decrease, and Forest Land area was the highest; under ED, the net changes in Forest Land were opposite in direction between the two bays. The future spatial expansion pathways of Construction Land also differed clearly between the two bays: Haitan Bay expanded mainly around existing built-up areas, whereas Tannan Bay showed a stronger transportation-corridor orientation.

(3)

Most land allocations remained consistent between ER and ED, while Construction-Expansion Divergence Areas mainly reflected allocation differences between retaining Cropland and Grassland and expanding Construction Land. Within Construction-Expansion Divergence Areas, NP, PD, and AREA_MN changed markedly with MMU settings, and the inter-bay rankings of NP, PD, and AREA_MN changed among MMU settings. Within Forest-Restoration Divergence Areas, the directions of inter-bay differences in AREA_MN and LPS remained consistent across MMU settings. Tannan Bay also had a greater extent of Forest-Restoration Divergence Areas and markedly higher Area Retention Rates under the two MMU filtering settings.