Integrating POI-Driven Functional Attractiveness into Cellular Automata for Urban Spatial Modeling: Case Study of Yan’an, China

Abstract

Urban growth models often prioritize environmental and accessibility factors while underestimating behavioral and functional dynamics. This study develops a POI-enhanced Cellular Automata (CA) framework to simulate urban expansion by incorporating three semantic indicators derived from Point-of-Interest (POI) data—density (PD), diversity (PDI), and functional centrality (FC). Taking Yan’an, China, as a case, the model integrates these indicators with terrain and infrastructure variables via logistic regression to estimate land-use transition probabilities. To ensure robustness, spatial block cross-validation was adopted to reduce spatial autocorrelation bias. Results show that the POI-based model outperforms the baseline in both Kappa and Figure of Merit metrics. High-density and mixed-function POI zones correspond with compact infill growth, while high-centrality zones predict decentralized expansion beyond administrative cores. These findings highlight how functional semantics sharpen spatial prediction and uncover latent behavioral demand. Policy implications include using POI-informed maps for adaptive zoning, ecological buffer protection, and growth hotspot management. The study contributes a transferable workflow for embedding behavioral logic into spatial simulation. However, limitations remain: the model relies on static POI data, omits vertical (3D) development, and lacks direct comparison with alternative models like Random Forest or SVM. Future research could explore dynamic POI trajectories, integrate 3D building forms, or adopt agent-based modeling for richer institutional representation. Overall, the approach enhances both the accuracy and interpretability of urban growth modeling, providing a flexible tool for planning in functionally evolving and ecologically constrained cities.

1. Introduction

Urbanization is one of the most transformative processes reshaping the global landscape, profoundly influencing land systems, environmental sustainability, and socio-economic interactions [1,2,3]. It is not merely a demographic or infrastructural phenomenon but a multi-scalar restructuring of space, society, and ecological relationships [4]. Over recent decades, urban expansion has intensified worldwide, particularly in developing countries, where population growth, industrialization, and modernization converge to produce rapid and often uncoordinated spatial growth [5,6,7,8]. This process has led to altered land-use configurations, fragmented urban forms, environmental pressures, and uneven service accessibility [9], posing challenges for sustainable development, ecological protection, and equitable resource distribution [10].

In this context, simulating and forecasting urban spatial evolution has become a core concern in both academic and policy spheres [11,12]. The concept of spatial governance—managing the distribution, connectivity, and transformation of land uses—has gained prominence in the pursuit of sustainable, efficient, and resilient cities [13]. Urban simulation models that integrate socio-economic behavior with spatial dynamics are increasingly regarded as indispensable for smart planning, land-use regulation, and environmental management [14,15,16]. With the combined pressures of population growth, climate change, and technological innovation, the need for models that are robust, interpretable, and empirically grounded has become more urgent [17].

Cellular Automata (CA) models are among the most widely used approaches for simulating land-use and urban development [18,19,20]. By representing space as a grid of cells whose states evolve through transition rules, CA captures spatial dependency and emergent patterns such as edge expansion, infill, and multi-nodal growth [21,22]. Yet, traditional CA models are often criticized for limited behavioral realism. Transition rules usually rely on static spatial features—such as distance to roads, centers, or water bodies—and past land-use patterns. While useful, such models struggle to capture the complex decision-making processes that underpin real urban development [23,24,25], including accessibility to services, functional attractiveness, market dynamics, and institutional constraints [26].

The proliferation of high-resolution spatial data has created new opportunities to enrich CA. Among these, Point-of-Interest (POI) data stand out for capturing the functional and behavioral dimensions of urban space [27,28]. POIs, representing facilities such as schools, hospitals, restaurants, parks, and government offices, are often extracted from digital maps, location-based services, or open databases [29,30]. Unlike remote sensing, which depicts surface conditions, POIs provide semantic information on how space is used and valued in daily life, serving as proxies for urban function, activity intensity, and locational preference.

Recent studies demonstrate the utility of POIs in urban morphology, functional zoning, and service accessibility [31,32,33]. They have been used to identify mixed-use centers, assess vitality, and evaluate land-use efficiency. Yet, despite these advances, POIs remain underutilized in predictive simulations. While effective for mapping patterns, they are rarely embedded in dynamic land-use change models. Given their behavioral content, POIs could significantly enhance the explanatory and predictive power of CA—particularly where human activity patterns shape urban development trajectories more strongly than physical constraints [34].

Embedding POI-based indicators such as density, diversity, and clustering into CA rules adds a functional dimension to simulation logic [35]. For instance, areas with dense and diverse POIs may exert stronger development pull, reflecting economic attractiveness and service accessibility. Such refinements enable CA to better reproduce observed growth patterns in mixed-use or rapidly changing contexts [36,37].

This study develops a POI-enhanced CA framework to simulate the evolution of urban spatial structure in a more behaviorally realistic way [38]. By integrating functional intensity and heterogeneity into transition rules, the model bridges the gap between physical modeling and behavioral urban analysis.

The framework is applied to the central urban area of Yan’an, a historically and environmentally significant city in northwestern China [39]. Located in the fragile Loess Plateau, Yan’an illustrates the tension between urban expansion, ecological conservation, and spatial governance [40]. The city also faces ecological constraints, including strict environmental regulations, erosion control, and urban growth boundaries defined by ecological redlines [41].

These dynamics make Yan’an a compelling case for testing behaviorally enhanced models [42]. Its development is shaped simultaneously by functional demand—such as tourism-driven expansion and infrastructure improvements—and by ecological and heritage-preservation mandates that constrain growth [43]. A POI-driven CA framework can identify functional pressures, simulate constrained expansion, and provide spatially nuanced forecasts for policy interventions.

Through this case, the study demonstrates how POI integration improves the accuracy, interpretability, and policy relevance of CA-based simulations. The results contribute both technical advances in urban modeling and practical insights for sustainable planning in ecologically and historically sensitive regions (Figure 1).

The central objectives of this study are multifaceted, reflecting both methodological innovation and practical applicability. First, it systematically integrates POI data into the CA framework in a way that captures not only the spatial intensity of human activities but also the functional diversity and semantic meaning embedded in urban space [31]. Rather than treating POIs as auxiliary references, the proposed framework operationalizes them as core drivers of land-use transition potential. This involves constructing functional intensity surfaces, diversity indices, and spatial clustering metrics from POI distributions, which are then embedded directly into CA transition rules to influence cell-level land-use change [31].

Second, the model is calibrated and validated using multi-temporal land-use datasets covering the period 2010–2024, a time of rapid transformation in Yan’an’s urban structure. This enables evaluation of the model’s ability to reproduce observed spatial expansion. Calibration adjusts transition probabilities to observed trajectories, while validation employs spatial accuracy measures and landscape metrics, ensuring empirical robustness [44,45].

Third, the study analyzes the relationship between simulated urban expansion and ecological protection zones—such as ecological redlines—and infrastructure development corridors, including roads and public transport [46]. Through spatial overlay analysis and conflict detection, the model assesses whether projected growth trajectories align with sustainable planning objectives or risk encroaching on environmentally sensitive or infrastructure-deficient areas [47].

Fourth, the framework supports scenario-based simulations, a critical feature for planning support systems. Three scenarios—business-as-usual, policy-constrained, and ecological-prioritized—are tested, with parameters such as functional attractiveness thresholds, resistance values in protected areas, and neighborhood weights adjusted to reflect different planning strategies. This allows comparison of alternative futures under varying institutional and environmental conditions [48,49].

From a methodological perspective, this research contributes by generating CA transition rules from POI-based indicators rather than conventional proximity metrics [50]. Indicators include POI density, diversity (e.g., Shannon entropy), centrality (e.g., weighted kernel density of key services), and spatial clustering (e.g., Getis–Ord Gi*), computed across multiple spatial scales [51]. In this way, the framework refines neighborhood influence mechanisms, enabling cells to respond to functional dynamics rather than generic adjacency [52].

Furthermore, the model enhances the behavioral realism of urban simulation by explicitly linking land-use change potential to accessibility and functional attractiveness [53]. This reflects the empirical reality that urban development is driven not only by physical proximity or administrative zoning but also by access to services, opportunities, and amenities [54].

From a planning and policy standpoint, the framework demonstrates strong applied value. POI-enhanced CA models can be embedded into geographic decision support systems to visualize the spatial implications of policy alternatives [55]. By simulating the effects of ecological constraints or service allocation strategies, planners can assess trade-offs between development intensity, spatial equity, and ecological protection. The model also aids in delineating growth boundaries, optimizing land-use mix, and identifying areas requiring intervention [56].

In sum, this study bridges the gap between data-rich urban analysis and behaviorally grounded simulation [57]. It demonstrates how integrating emerging spatial datasets with established modeling frameworks can advance both theoretical understanding and practical planning for sustainable urban development [58].

The remainder of this paper is structured as follows: Section 2 reviews literature on POI applications and CA-based urban simulation; Section 3 presents the methodological framework; Section 4 introduces the study area and datasets; Section 5 reports simulation results and scenario analyses; Section 6 discusses implications for governance, ecology, and infrastructure; and Section 7 concludes with theoretical contributions, methodological innovations, and policy relevance, while outlining future directions such as cross-city transferability and dynamic POI monitoring (Figure 2).

2. Literature Review

Urban simulation has increasingly become a central theme in spatial planning and urban science, driven by the accelerated pace of urbanization, the need for environmental sustainability, and the rising complexity of spatial systems [59]. In the face of global climate pressures, demographic shifts, and socio-economic transformation, cities have evolved into highly dynamic entities whose spatial expansion patterns defy simple deterministic models [60]. As urban environments grow in both horizontal scale and vertical intensity, the capacity to accurately model, analyze, and predict their development trajectories becomes a pressing concern for spatial governance [61]. Urban simulation, therefore, represents not only a methodological frontier but also a strategic instrument for land-use optimization, infrastructure investment, environmental risk mitigation, and long-term spatial planning [62].

Traditional modeling approaches have often fallen short in capturing the nuanced and multi-dimensional nature of urban growth. Static land suitability analyses, simple regression models, and top-down zoning frameworks have proven insufficient to handle the emergent properties of urban systems—such as feedback loops, non-linearity, and spatial heterogeneity [63]. In response, scholars and practitioners have increasingly turned to dynamic, rule-based, and behaviorally inspired models. Among these, CA models have garnered particular attention for their ability to simulate spatial change processes over time, and POI data have emerged as a valuable resource to characterize the socio-functional composition of urban space [64]. The convergence of these two domains—CA simulation and POI analytics—offers a promising yet underdeveloped path toward enhancing the behavioral realism and spatial resolution of urban modeling [65].

CA models, initially developed in mathematical and biological contexts, have been adapted to urban studies since the 1990s as tools to simulate land-use transitions and urban form evolution [66]. The foundational principle of CA lies in its discrete, cell-based representation of space, wherein each cell can change its state based on predefined rules that consider the states of neighboring cells. In urban applications, CA models are typically used to simulate how undeveloped or agricultural land converts into residential, commercial, or industrial uses. These transitions are influenced by a variety of spatial determinants, such as distance to roads, existing built-up areas, topographic conditions, and planning constraints [67]. The spatial-temporal logic of CA makes it well-suited for projecting the diffusion of urbanization across landscapes, especially when combined with geographic information systems (GIS) and multi-source spatial datasets [68].

Over time, the CA framework has been expanded to include a range of improvements. These include the use of multi-criteria evaluation for transition potential, stochastic disturbance to reflect random variability, and hybridization with artificial intelligence techniques such as logistic regression, neural networks, and fuzzy logic [69]. Spatial calibration has also advanced, with techniques such as Markov chains, Monte Carlo simulations, and genetic algorithms being employed to fine-tune transition rules [70]. Furthermore, multi-scale and hierarchical CA models have been developed to accommodate complex urban systems operating at neighborhood, district, and metropolitan scales simultaneously [71].

Beyond raster-based models, scholars have also developed Vector Cellular Automata (V-CA), which simulate land-use transitions on irregular units such as cadastral parcels or planning zones [72]. V-CA has demonstrated advantages in incorporating zoning regulations, floor-area ratios, and parcel-specific constraints, thereby improving the realism of policy-sensitive simulations [73]. Studies have shown that parcel-level CA can better represent institutional drivers and fine-grained urban morphology [74]. Nevertheless, raster-based CA remains widely used due to its scalability, data availability, and ease of cross-city transfer. This study, therefore, adopts a raster CA framework but enriches it with POI-based functional indicators, aiming to combine the scalability of raster CA with the behavioral sensitivity typically associated with vector approaches. Despite these methodological innovations, one core limitation of CA remains: its limited behavioral depth [75]. While CA is effective at modeling morphological expansion, it often lacks the ability to reflect human-centered drivers of urban change.

This behavioral deficiency stems from the overreliance on physical or infrastructural variables—e.g., road networks, slope, land-use adjacency—that only partially capture the decision-making logic of individuals, businesses, or institutions involved in urban development [76]. In most conventional CA applications, transition rules are generated from spatial proximity metrics or static environmental layers. Although such models can replicate the general contours of urban sprawl or infill development, they fall short in accounting for the socio-economic attractiveness of different locations, the role of accessibility to services, or the clustering tendencies of specific land-use types [77]. As a result, the spatial outputs may be formally accurate but behaviorally naive, limiting their utility for planners and policymakers seeking to intervene in real-world development processes [78].

To overcome this gap, scholars have sought to incorporate richer data sources that capture the functional and behavioral logics underlying spatial change. One of the most promising developments in this direction is the integration of geospatial big data—particularly POI data—into simulation frameworks [79]. POI data, typically harvested from digital mapping platforms, location-based services, and volunteered geographic information, provide detailed information about the location, type, and density of urban functions such as retail, education, healthcare, finance, and leisure [32]. Unlike satellite imagery or land-cover classifications, which depict physical forms, POI datasets offer insight into the operational content of urban space—how it is used, by whom, and for what purposes [80].

In recent years, POI data have been widely applied in urban analysis to reveal functional structures, activity clusters, and service distributions [31]. Researchers have employed POI clustering to delineate centers and mixed-use zones, while POI-derived indicators have informed assessments of accessibility, urban vitality, and spatial equity [81]. These applications highlight POIs’ advantage in capturing the socio-functional dimension of cities—the flows of people, services, and opportunities that shape urban life [82].

However, the vast majority of POI-related studies remain descriptive or diagnostic in nature. They map patterns, visualize clusters, or assess disparities, but rarely engage with dynamic modeling or predictive simulation [83]. This reflects a broader divide in urban science: the disjunction between spatial diagnosis and process modeling. POI data, rich in behavioral signals, have yet to be systematically incorporated into simulation models that can anticipate change, test planning scenarios, or evaluate policy impacts [84]. This is particularly true for CA-based models, which have traditionally been parameterized using static spatial features and have not fully capitalized on the potential of behaviorally meaningful datasets.

Integrating POI data into CA models offers a pathway to bridge this divide. Theoretically, POI distributions provide proxies for urban demand, locational preference, and functional hierarchy—all of which are critical for determining the probability of land conversion [85]. Practically, embedding POI-based variables into CA transition rules allows the model to “sense” the socio-functional context of each cell, thereby enhancing its behavioral realism. Empirically, POI data are often updated in near real time and cover a wide range of urban functions, making them suitable for dynamic simulation workflows [86]. Methodologically, they enable the development of hybrid models where physical proximity and functional attractiveness jointly determine land-use change [87].

Some recent studies have begun to explore this integration. For example, POI density has been used as an explanatory variable in logistic regression models that inform CA transition probabilities [88]. Others have employed clustering algorithms (e.g., DBSCAN, k-means) to define urban functional zones from POI distributions, which then serve as decision constraints or suitability modifiers in CA models [89]. These experiments have demonstrated that the inclusion of POI-derived indicators can improve the accuracy, resolution, and interpretability of urban growth simulations, especially in transitional zones where traditional proximity metrics are insufficient [90]. Nevertheless, such efforts remain limited in scale, fragmented in methodology, and largely ad hoc in their implementation [91].

Several critical gaps remain. First, few studies systematically explore the differential effects of various POI categories—such as retail, education, or public services—on land-use change. Second, most existing models treat POI variables as static layers, failing to incorporate their temporal evolution or volatility. Third, there is little effort to develop generalized, transferable frameworks for POI-CA integration that could be applied across different urban contexts. Finally, the theoretical grounding of such integrations often remains underdeveloped, lacking a clear articulation of how POI data relate to urban development theory, spatial choice behavior, or policy drivers [92].

Addressing these limitations calls for a more integrated, theoretically informed, and empirically validated approach. Specifically, there is a need to develop CA simulation models that embed POI-based indicators as endogenous drivers of spatial change, rather than as exogenous constraints or background layers. Such models should account for POI diversity (entropy), clustering intensity (spatial autocorrelation), functional proximity, and temporal shifts, and relate these to land development probabilities. Moreover, they should support scenario-based analysis, enabling planners to simulate different development trajectories under varying assumptions of infrastructure provision, economic growth, or zoning reform [93].

The potential benefits of such integration are manifold. POI-enhanced CA models can capture latent development pressures, detect emerging centers of activity, and simulate fine-grained changes in urban morphology. They also facilitate multi-dimensional policy evaluation—assessing not just where land use will change, but why, for whom, and with what socio-economic implications. This is particularly critical in cities that are simultaneously managing growth, inequality, and ecological risk [94]. Medium-sized cities with constrained land resources and significant planning pressures, such as Yan’an, offer ideal testbeds for such models. In these contexts, the integration of POI data into CA models can inform balanced development, guide infrastructure investment, and support adaptive spatial strategies [95].

The present study responds to this challenge by developing a novel POI-driven CA modeling framework for simulating urban spatial structure in data-sparse and ecologically sensitive environments. Unlike prior efforts that treat POI variables as supplementary inputs, this framework integrates them directly into the CA transition logic. Transition probabilities are modified based on POI density gradients, category-specific attractiveness scores, and functional clustering coefficients. The model is empirically calibrated using multi-period land-use data and POI datasets for Yan’an’s central urban area, allowing for validation against historical expansion patterns and testing under different planning scenarios. Special attention is given to polycentric development patterns, functional hierarchy, and spatial spillovers, reflecting the complex reality of contemporary urbanization [96].

In conclusion, while CA modeling and POI analytics have both matured as independent research strands, their integration remains at a preliminary stage. The literature has advanced substantially in refining spatial simulation techniques and in extracting behavioral insights from big data, but has yet to converge on robust methodologies that combine the strengths of both. This review has outlined the historical evolution, methodological advances, limitations, and future prospects of CA and POI research, highlighting the potential of their convergence to produce richer, more actionable models of urban spatial change [97]. By proposing and demonstrating an integrated POI-CA framework, the present study contributes to a new generation of urban simulation models that are both spatially grounded and behaviorally informed (

Table 1. Comparison between Traditional CA Models and POI-Enhanced CA Models in Urban Simulation.

Dimension	Traditional CA Models	POI-Enhanced CA Models
Main Data Inputs	Physical variables: distance to roads, land suitability, elevation, slope	Functional variables: POI density, diversity, centrality + physical + policy constraints
Behavioral Representation	Limited; often static and proximity-based	Improved; reflects functional attractiveness and urban activity intensity
Functional Sensitivity	Low: ignores real-world functional structure	High: sensitive to localized socio-economic behavior and service agglomeration
Spatial Resolution	Often coarse-grained, less reflective of intra-urban heterogeneity	Fine-grained; captures micro-scale functional patterns through POI indicators
Transition Rule Mechanism	Predefined or regression-based on physical variables only	Data-driven; logistic regression with POI-based indicators integrated
Predictive Capacity	Good for morphological patterns but poor in identifying function-driven growth	Improved ability to predict growth in active functional nodes and emerging centers

).

3. Methodology

This study proposes a simulation framework that integrates POI data into a CA model to improve the behavioral and functional sensitivity of urban spatial structure prediction. The overall approach is organized into five components: definition of the study area and spatial units, data acquisition and preprocessing, estimation of land-use transition probabilities using POI-based indicators, implementation of the CA simulation process, and scenario analysis. The design rationale and theoretical basis for each component are elaborated in the subsections below.

3.1. Data Acquisition and Preprocessing

This study integrates multi-source spatial data to capture both the physical and functional characteristics of the urban landscape. All datasets were resampled to a uniform 30 m resolution raster grid to ensure variable comparability and compatibility with cellular automata (CA) modeling. The input variables fall into four main categories: land use, POI-derived functional indicators, transportation accessibility, and topographic constraints.

Land-use maps for the years 2010, 2015, 2020, and 2024 were generated through supervised classification of medium-resolution satellite imagery, validated against official land survey data. Each pixel was assigned to one of six categories: built-up land, cultivated land, forest, grassland, water body, and unused land. These data serve both as the historical input for training and as reference targets for simulation validation.

Point-of-Interest (POI) data were collected from the Gaode (Amap) open API and classified into six major functional types: residential, commercial, educational, medical, governmental, and recreational. Rather than using raw POI counts, we constructed three derived indicators for each grid cell:

POI Density (PD): the number of POIs within a 300 m circular buffer, capturing activity intensity at the neighborhood scale.

POI Diversity Index (PDI): calculated using Shannon entropy to measure functional mixture:

$${PDI}_i=-\sum _{j=1}^m{p}_{ij}ln\left(p_{ij}\right)$$

where $p_{ij}$ is the proportion of POIs of category j in cell i, and m = 6 is the number of POI types.

Functional Centrality (FC): kernel density estimation of higher-order services (e.g., hospitals, government offices, shopping centers), computed using a Gaussian kernel with a 400 m radius and a bandwidth (σ) of 100 m to reflect decaying functional influence with distance.

These indicators were selected to represent three complementary dimensions of urban spatial behavior: intensity (PD), heterogeneity (PDI), and dominance (FC). Compared to alternative socio-economic indicators such as population density or land prices, POIs provide higher spatial granularity, semantic richness, and open accessibility. In Chinese cities where dynamic micro-scale data are often unavailable, POI-based proxies offer a stable and reproducible representation of urban functional structure.

To verify their spatial relevance, we conducted a bivariate correlation analysis between the 2020 built-up land and each POI indicator. Results showed strong spatial associations: PD (r = 0.64), PDI (r = 0.51), and FC (r = 0.69), all statistically significant at p < 0.001. These findings confirm that POI distributions are not merely descriptive but carry meaningful predictive signals for land-use transformation (

Table 2. Correlation between POI indicators and 2020 built-up land distribution.

Indicator	Pearson Correlation	Significance
POI Density (PD)	0.64	p < 0.001
POI Diversity (PDI)	0.51	p < 0.001
Functional Centrality (FC)	0.69	p < 0.001

).

Additionally, sensitivity tests on buffer radius and kernel bandwidth were conducted. For PD and FC, values ranging from 200~500 m were evaluated. The resulting spatial patterns remained stable (standard deviation of built-up probability < 0.06), demonstrating the robustness of parameter selection.

Transportation accessibility was derived from OpenStreetMap road networks. For each cell, the Euclidean distance to the nearest primary or secondary road was calculated as a measure of connectivity. Topographic constraints were captured using slope and elevation data extracted from the ASTER Global DEM. These variables reflect natural barriers and cost surfaces that influence urban expansion.

To facilitate integration into the CA model, all continuous variables (PD, PDI, FC, distance to road, slope, elevation) were normalized to the range [0, 1] using min–max scaling. This ensures comparability and prevents bias toward variables with larger numeric ranges.

3.2. Transition Probability Estimation Using POI Indicators

To estimate the probability of urban land conversion, we employed a binary logistic regression model. This method was selected due to its transparency, probabilistic formulation, and widespread use in urban modeling, particularly for studies seeking to interpret the influence of explanatory variables on land-use transitions.

Let $P_i$ denote the probability that grid cell i transitions from non–built-up to built-up land between 2015 and 2020. The logistic regression model is specified as:

$$P_i={\displaystyle \frac{1}{1+e^{-\left({\beta }_0+\sum _{k=1}^n{\beta }_k{X}_{ik}\right)}}}$$

where X_ik represents the k-th explanatory variable for cell i, and ${\beta }_k$ are coefficients estimated via maximum likelihood. The dependent variable is binary: 1 if transition occurred, 0 otherwise. The model includes both traditional spatial predictors and functional indicators derived from POI data (

Table 3. Descriptions of explanatory variables used in the logistic regression model.

Variable	Description
Distance to road	Euclidean distance to nearest primary/secondary road
Slope	From ASTER GDEM
Elevation	From ASTER GDEM
POI Density (PD)	Activity intensity
POI Diversity Index (PDI)	Functional mixture (Shannon entropy)
Functional Centrality (FC)	Kernel density of higher-order POIs
Composite Development Index (CDI)	Weighted sum: CDI = 0.4 × PD + 0.3 × PDI + 0.3 × FC (used for robustness check only)

).

The model was trained on 10,000 stratified random samples (5000 transition cells and 5000 stable cells) to ensure class balance. Coefficient estimates were tested for significance, and all variables were statistically significant at the 0.001 level. To assess multicollinearity, Variance Inflation Factors (VIFs) were calculated. All VIF values were below 2.5, indicating low collinearity and stable estimation (

Table 4. Logistic regression results for land-use transition probability.

Variable	Coefficient (β)	Std. Error	z-Value	p-Value	VIF
Intercept	−1.841	0.093	−19.8	<0.001	—
Distance to road	−0.457	0.026	−17.6	<0.001	1.32
Slope	−0.261	0.021	−12.4	<0.001	1.24
Elevation	−0.142	0.019	−7.5	<0.001	1.18
POI Density (PD)	0.782	0.037	21.1	<0.001	2.04
POI Diversity (PDI)	0.619	0.034	18.1	<0.001	1.97
Functional Centrality (FC)	0.835	0.042	19.9	<0.001	2.18

).

To assess the statistical robustness of the logistic regression model, we computed standard errors, z-values, and p-values for each estimated coefficient. As shown in Table 4, all POI-related variables—density (PD), diversity (PDI), and functional centrality (FC)—exhibited high statistical significance (p < 0.001), confirming their strong predictive power in urban land-use transition.

Additionally, we examined potential multicollinearity among the POI indicators using both Pearson correlation coefficients and the Variance Inflation Factor (VIF). The highest pairwise correlation was 0.51 between POI density and diversity, and all VIF values were below 2.1, well within acceptable limits. These results suggest that the POI variables capture distinct yet complementary functional dimensions and do not suffer from problematic collinearity.

To minimize overfitting and account for spatial autocorrelation, we implemented a spatial block cross-validation strategy using the “checkerboard” method. The study area was divided into cell blocks, with 5-fold cross-validation applied: in each fold, 80% of blocks were used for training and 20% for validation. The model yielded an average AUC of 0.842 ± 0.015 and average accuracy of 0.796 ± 0.021 across folds, indicating strong generalization performance across independent spatial partitions (

Table 5. Spatial cross-validation results.

Fold	AUC	Accuracy
1	0.845	0.791
2	0.837	0.794
3	0.851	0.812
4	0.826	0.778
5	0.850	0.805
Mean ± SD	0.842 ± 0.015	0.796 ± 0.021

).

Although advanced machine learning models such as Random Forest (RF), Gradient Boosting (GBDT), or Support Vector Machines (SVM) may yield higher predictive accuracy, we prioritized interpretability and ease of integration into the CA simulation framework. The logistic regression model enables explicit explanation of functional drivers and their behavioral influence on urban expansion.

Future studies may include benchmark tests comparing logit with RF or GBDT to quantify performance gains. At present, we acknowledge this limitation and highlight the importance of interpretable predictors for theory-driven simulation.

3.3. Cellular Automata-Based Simulation of Urban Growth

Following the estimation of transition probabilities, the spatial evolution of urban land was simulated using a Cellular Automata (CA) model. CA is particularly effective for capturing bottom-up urban growth processes, as it incorporates spatial dependencies, local interactions, and environmental constraints within a rule-based framework. In this study, each grid cell represents a discrete land-use state—either built-up or non–built-up—and updates iteratively over time according to a transition rule that integrates probabilistic, neighborhood, and constraint factors.

The CA transition rule for each cell i is defined as follows:

$$T_i=P_i\times \left(1+a{N}_i\right)\times C_i$$

where

• T_i: Final transition score of cell i;
• P_i: Transition probability from logistic regression;
• N_i: Neighborhood effect, defined as the proportion of urban cells in the 3 × 3 Moore neighborhood of cell i;
• a: Neighborhood weight coefficient, empirically calibrated to balance endogenous expansion and exogenous drivers (set to 0.5 in this study based on validation);
• C_i: Binary constraint mask (1 if the cell is developable, 0 if restricted).

The Moore neighborhood configuration was adopted in this study, wherein each cell interacts with its eight immediate neighbors. This configuration is better suited to capturing the realistic patterns of urban growth, as it allows for multidirectional expansion, lateral diffusion, and leapfrog development. The neighborhood effect coefficient (denoted as a) was empirically calibrated using observed urban expansion data from 2010 to 2015. The coefficient was set at 0.5 to reflect a balanced representation of spatial contagion, where built-up cells tend to emerge in proximity to existing urban areas, yet still allowing for peripheral developments.

To ensure that the simulation adhered to both environmental and policy-based restrictions, a comprehensive constraint mask was constructed and applied to each simulation iteration. This binary raster layer was composed of three types of constraints. First, physical constraints included natural features such as water bodies, slopes steeper than 25 degrees, and ecologically protected areas identified from topographic and conservation datasets. Second, policy constraints incorporated urban growth boundaries (UGBs) and ecological redlines, digitized from the Yan’an Master Plan and provincial land-use control maps. Third, suitability filtering was applied whereby cells with extremely low development potential (logistic regression probability below 0.1) were excluded to maintain computational efficiency and policy realism. Together, these constraints formed a mask that strictly governed where development could or could not occur, thus enforcing legal and ecological boundaries throughout the simulation process.

The simulation was initialized with the 2015 land-use map and executed in annual iterations until 2024. During each iteration, a three-step procedure was followed. First, the number of cells to be converted to built-up land was estimated based on the historical urban expansion rate (2015–2020 average). Second, all non-built-up and unconstrained cells were ranked by their final transition scores, which incorporated both logistic regression probabilities and neighborhood effects. Finally, the top-ranked cells were selected and converted to built-up land, ensuring that the total area of new development per year matched the calibrated annual growth rate. This temporal disaggregation better reflects real-world planning cycles, construction phases, and land market rhythms than a single-step static simulation.

While the Cellular Automata framework used in this study effectively reproduces the spatial diffusion of urban land through horizontal growth mechanisms, it remains limited in several respects. Vertical urbanization processes, such as building densification, changes in floor-area ratio (FAR), or vertical land-use mixing, are not captured. Additionally, institutional dynamics—including land lease policy changes, developer negotiations, or bottom-up community resistance—are not explicitly modeled, although they are partially represented through constraint layers. The model also lacks stochastic components, as it employs a deterministic annual allocation strategy without random perturbations or adaptive learning mechanisms.

Future research could address these limitations through several avenues. One possibility is to integrate agent-based modeling to simulate interactions between planning institutions, developers, and residents. Another is to incorporate 3D spatial modeling frameworks, allowing for simultaneous simulation of both horizontal expansion and vertical intensification. Lastly, the application of stochastic Cellular Automata with Monte Carlo ensemble runs would allow for uncertainty quantification and probabilistic forecasting, thereby enhancing the robustness and policy utility of the simulation outcomes (

Table 6. Cellular Automata Simulation Parameters and Configuration.

Parameter	Description
Neighborhood type	Moore neighborhood (3 × 3), includes 8 adjacent cells
Neighborhood weight coefficient	a = 0.5, calibrated using 2010–2015 empirical expansion
Constraint types	Physical (water, slope > 25°, ecological zones); policy (UGB, redlines); low suitability filter (p < 0.1)
Transition probability	Derived from logistic regression based on spatial and functional indicators
Annual allocation unit	Based on average annual expansion between 2015 and 2020
Temporal iteration	Year-by-year update from 2015 to 2024
Initialization year	2015 land-use map
Allocation strategy	Rank-based conversion of top-scoring unconstrained cells per year
Limitation scope	No modeling of FAR, land policy feedback, or stochastic behavior

).

3.4. Model Calibration and Accuracy Assessment

To ensure the reliability and interpretability of the POI-driven CA simulation framework, a two-step validation procedure was implemented: model calibration using observed transitions (2015–2020), followed by accuracy assessment via cross-validation and spatial performance metrics.

The CA model was calibrated to replicate the spatial and temporal dynamics of observed land-use change between 2015 and 2020. Key parameters adjusted include:

Neighborhood weight coefficient (α): controlling the influence of nearby built-up cells on transition probability;

Annual conversion quota: reflecting the number of new urban cells per year, based on historical trends.

Calibration aimed to minimize visual and statistical discrepancies between simulated and observed expansion patterns during this baseline period.

To reduce overfitting and address spatial autocorrelation, a spatial block cross-validation approach was applied. The study area was divided into contiguous blocks, and model training and validation were conducted on geographically independent subsets. This method enhances statistical robustness by avoiding leakage from spatially proximate samples.

Three widely used metrics were employed to assess accuracy:

• Overall Accuracy (OA): percentage of correctly predicted cells;
• Kappa Coefficient ($k$): agreement beyond chance;
• Figure of Merit (FoM): overlap of correctly predicted transitions with observed ones.

At the core of the validation process is the comparison between the simulated 2020 land-use map L^sim2020 and the observed map L^obs2020. Three evaluation metrics were employed:

Overall Accuracy (OA):

$$\mathrm{O}\mathrm{A}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}}}$$

where TP is the number of correctly predicted built-up cells, TN is the number of correctly predicted non-built-up cells, FP and FN are the false positives and false negatives, respectively. OA reflects the proportion of correctly predicted outcomes across all land types.

Kappa Coefficient ($k$):

$$k={\displaystyle \frac{P_o-P_e}{1-P_e}}$$

where P_o is the observed accuracy (same as OA), and P_e is the expected accuracy by random chance. Kappa accounts for random agreement and is a more conservative indicator of model reliability.

Figure of Merit (FoM):

$$\mathrm{F}\mathrm{o}\mathrm{M}={\displaystyle \frac{\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a} \mathrm{c}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{l}\mathrm{y} \mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{s} \mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}}{\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a} \mathrm{o}\mathrm{b}\mathrm{s}\mathrm{e}\mathrm{r}\mathrm{v}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{s} \mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}+\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a} \mathrm{s}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{s} \mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}-\mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a} \mathrm{c}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{l}\mathrm{y} \mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{s} \mathrm{c}\mathrm{h}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}}}$$

The POI-enhanced CA achieved high accuracy (OA ≈ 91%, \kappa = 0.82, FoM = 0.64). However, these values should be interpreted with caution, as spatial dependence and model assumptions can inflate accuracy metrics. The adoption of spatial block validation mitigates this risk, but further benchmarking against alternative classifiers (e.g., random forest, gradient boosting) would provide additional robustness checks.

In addition to quantitative validation, visual comparison of simulated versus observed maps allowed for qualitative assessment. The model successfully reproduced major expansion corridors and growth hotspots, while also capturing emerging infill nodes within the urban core. Comparative experiments without POI variables confirmed that the inclusion of POI-based indicators improved both the spatial resolution and behavioral interpretability of predicted patterns.

Finally, sensitivity analysis was conducted on the neighborhood weight parameter (\alpha) that controls the relative influence of neighboring built-up cells. Higher values of \alpha led to more clustered expansion patterns, while lower values encouraged more scattered development. By systematically adjusting this parameter, the model reproduced Yan’an’s observed tendency toward clustered growth along infrastructure corridors, thereby increasing confidence in the robustness of the simulation outcomes.

3.5. Scenario Simulation and Policy Implications

To assess the practical utility of the POI-enhanced CA framework for planning decision-making, this study constructed two contrasting development scenarios covering the 2015–2024 horizon. These scenarios were designed to reflect divergent assumptions regarding ecological constraints and spatial governance priorities. By juxtaposing unconstrained urban growth against policy-restricted development, the simulations allow for a nuanced understanding of trade-offs in spatial form, ecological integrity, and land-use intensity.

In the Business-as-Usual (BAU) scenario, urban expansion unfolds according to endogenous spatial dynamics. Transition probabilities derived from logistic regression interact with neighborhood diffusion effects and terrain-based exclusions to simulate growth. No additional ecological or regulatory constraints are applied. As a result, simulated development tends to follow corridors of low resistance—typically along existing infrastructure, flat terrain, and areas of high functional accessibility. The urban form under this scenario displays a dispersed pattern of peripheral expansion, with limited consolidation in existing urban cores.

In contrast, the Ecological Constraint (ECO) scenario enforces a stricter regulatory environment. A composite exclusion mask was constructed by integrating ecological redlines, urban growth boundaries (UGBs), and slope-sensitive development bans. These constraints were digitized from local planning documents and encoded into the simulation to restrict land conversion in environmentally or institutionally protected areas. The application of these constraints results in a more compact growth morphology, favoring inward expansion and redevelopment of underutilized land parcels within the core urban fabric. Compared to BAU, the ECO scenario substantially reduces outward sprawl and preserves sensitive ecological zones.

The key differences between the two scenarios are summarized in

Table 7. Comparative Characteristics of BAU and ECO Urban Growth Scenarios.

Aspect	BAU Scenario	ECO Scenario
Constraint Mechanism	Terrain-based only (water bodies, steep slopes)	Terrain + policy constraints (UGBs, redlines, slope ≥ 25°)
Functional Drivers	Logistic regression + POI influence	Same as BAU
Expansion Morphology	Peripheral spread, corridor-following, low-density clusters	Compact form, infill intensification, containment within cores
Ecological Zone Impact	Potential encroachment	Strict exclusion
Growth Control Intensity	Minimal	High
Development Hotspots	Along infrastructure and POI-rich peripheries	Reallocated to inner urban zones

, which highlights simulation inputs, constraints, and resulting urban morphology.

Despite their contrasting outputs, both scenarios reveal conflict-prone zones—areas where high development potential (indicated by POI density, diversity, and centrality) intersects with regulatory exclusion. These zones, often located at the urban-rural fringe, reflect the spatial tension between functional demand and conservation mandates. Their identification offers actionable insight for planners seeking to reconcile ecological protection with economic development. In particular, such zones may be candidates for conditional development, adaptive zoning, or policy negotiation.

The scenario simulation also underscores the value of embedding POI-derived indicators into urban growth models. Rather than simulating physical suitability alone, the inclusion of functional variables enables a behavioral lens, identifying not only where development can occur but also where it is most likely to concentrate based on spatial activity patterns. This functional realism strengthens the relevance of the model for strategic planning tasks such as urban containment policy design, redevelopment prioritization, and spatial equity assessments.

Furthermore, the modeling framework exhibits strong extensibility. Additional policy dimensions—such as floor area ratio incentives, public transit investments, or fiscal subsidies—could be integrated to simulate more complex interactions between land-use decisions and regulatory mechanisms. The inclusion of dynamic POI streams (e.g., real-time mobility data, service platform API feeds) offers a promising avenue for forecasting demand-responsive urban dynamics under shifting socio-economic conditions.

Overall, the scenario analysis demonstrates the ability of the POI-enhanced CA model to serve as a forward-looking planning tool. By simulating diverse futures and highlighting areas of strategic concern, it provides evidence-based support for balancing urban growth with environmental and institutional constraints.

4. Study Area and Dataset

4.1. Study Area and Regional Context

The central urban area of Yan’an, located in northern Shaanxi Province between the Ordos Plateau and the Loess Plateau, serves as the focus of this study (Figure 3). About 400 km north of Xi’an, Yan’an occupies over 37,000 km², though this research concentrates on its urban core and development fringe of roughly 300 km².

Ecologically, Yan’an lies in one of China’s most fragile environments. The Loess Plateau is marked by gullies, steep slopes, and high erosion risk, which constrain construction and necessitate ecological redlines, slope thresholds, and urban growth boundaries (UGBs) (Figure 4). Despite these restrictions, Yan’an has expanded notably over the past two decades, supported by government investment, industrial development, and the rise of service and tourism sectors. This growth has created tensions between functional intensification—particularly service- and tourism-related activities—and ecological or regulatory constraints.

Figure 4 uses NDVI, population density, precipitation, ecological values, and GDP density data for the entire Loess Plateau region. Yan’an is located in the better ecological-economic region.

The defined study area spans approximately 300 km² within Baota District and includes the historical river-valley core—home to government institutions, cultural landmarks, and dense service clusters—as well as newly developed commercial strips, large residential communities, and emerging sub-centers along transport corridors (Figure 5). The area also incorporates transitional urban–rural fringes, where informal housing and rural settlements coexist with new development zones. These interfaces are the most dynamic frontiers of land-use change, characterized by blurred functional boundaries, inconsistent planning implementation, and strong variability in POI distributions.

The study area shown in Figure 5 is the central urban area of Yan’an City, which is located in Baota District, with a spatial area of about 294.57 km², covering from Xishilipu on the east bank of the Yanhe River in the east to Nanqiaogou and Fenghuangshan in the west, Ganguyi Street in the south, and Wangjiaping Street in the north. Based on the Yan’an City Master Plan (2020–2035) and the Ministry of Natural Resources (MNR) Urban Boundary Delimitation Data 2021, the study area is composed of the “central city construction land boundary + peripheral transition area”, which is confirmed by using remotely sensed data superimposed on the administrative map layer. The geographic coordinates of the center of the study area are approximately 109°30′ E, 36°35′ N, with a geomorphological structure of “north–south trending river valley belt & hilly extension belt”. This boundary effectively captures both the formally planned urban construction area and the actual fringe of land conversion, serving as a representative zone for exploring planning-implementation gaps.

The study area boundary was delineated using the Yan’an City Master Plan (2020–2035) and Ministry of Natural Resources (MNR) urban boundary data (2021), cross-verified with remote sensing. The resulting simulation grid (30 m resolution) contains about 214,000 cells, each storing attributes such as land-use type, POI-derived functional indicators, slope, elevation, and accessibility to roads.

This integrated context—ecological fragility, heritage-driven service clustering, and policy-layered governance—makes Yan’an an ideal testbed for POI-enhanced Cellular Automata modeling. The selected scope captures both dense urban cores and transitional fringes, providing analytical richness for assessing how physical constraints and behavioral drivers jointly shape urban spatial structure.

4.2. Data Sources and Limitations

To support a behaviorally responsive urban growth simulation, this study constructed an integrated spatial dataset incorporating morphological features, functional signals, and planning constraints, all harmonized onto a consistent 30 m resolution raster grid. The land-use data, derived from high-resolution satellite imagery and classified into six standard categories, cover the years 2010, 2015, and 2020, with a projected layer for 2024 generated to facilitate near-future validation. All land-use datasets were processed to ensure categorical consistency, geospatial alignment, and temporal comparability, serving as both historical baselines and evaluation benchmarks.

The POI data were acquired from Gaode’s open platform in 2023, representing the most recent functional landscape of Yan’an’s urban activity patterns. Rather than being treated as raw spatial events, the POI dataset was systematically cleaned, filtered, and mapped onto the simulation grid as latent indicators of service accessibility and urban functional clustering, which reflect how people interact with space and influence development pressures. While POIs do not capture physical morphology, they reflect the distributional logic of urban service functions and spatial demand, which underpins the behavioral transition mechanisms embedded in the simulation model.

In addition to functional and land-use data, topographic inputs—including elevation and slope—were extracted from ASTER GDEM and resampled to match the simulation resolution. Transportation data were obtained from OpenStreetMap, focusing on primary road infrastructure relevant to accessibility modeling. Planning-related spatial constraints, such as urban growth boundaries (UGBs) and ecological redlines, were digitized from official planning documents and validated against the Yan’an City Master Plan and Ministry of Natural Resources boundary datasets.

Despite careful preprocessing, several limitations remain. The POI dataset captures only a static temporal snapshot, limiting its responsiveness to dynamic land-use transformations over time. Although spatial normalization and entropy-based modeling partially mitigate this limitation, temporal asynchrony between POIs and land-use data cannot be fully eliminated. Furthermore, while satellite-derived land-use classification achieves overall accuracies exceeding 90%, misclassifications in dense or topographically complex zones may persist. Topographic and accessibility variables may also be subject to positional uncertainties due to resolution constraints or source heterogeneity.

These limitations notwithstanding, the assembled dataset provides a robust foundation for behaviorally informed simulation. By integrating diverse spatial layers—ranging from remotely sensed morphology to functional clustering and regulatory boundaries—the modeling environment achieves both structural completeness and analytical tractability, enabling the exploration of complex urban growth scenarios with empirical grounding (

Table 8. Summary of Spatial Datasets Used in the Model.

Dataset Type	Source	Temporal Coverage	Spatial Resolution	Purpose
Land-Use Maps	Remote sensing (GF-2)	2010, 2015, 2020	30 m	Simulation base layers & validation
POI Data	Gaode API	2023	30 m grid	Functional behavior indicators
Road Network	OpenStreetMap	2022	Vector	Accessibility computation
DEM & Slope	ASTER GDEM	2011	30 m	Development constraints
Planning Boundaries	Yan’an Master Plan	2020–2035	Vector	UGBs and ecological redlines enforcement

). This integrated spatial database is not only technically comprehensive but also problem-driven, allowing for the exploration of functional-ecological trade-offs and policy-responsive urban growth trajectories in Yan’an’s unique planning context.

5. Results

5.1. Simulation Performance and Spatial Accuracy Assessment

This section evaluates the performance of the POI-enhanced Cellular Automata (CA) model in reproducing urban land-use transitions in Yan’an from 2010 to 2024. Model accuracy was assessed from three dimensions: spatial agreement with empirical land-use data, behavioral plausibility of growth patterns, and robustness of functional integration.

To ensure methodological transparency and minimize overfitting risks, model validation followed a temporal holdout strategy. The logistic regression component was trained using observed land-use transitions from 2010 to 2015, while validation was conducted against the 2020 land-use configuration. This approach allows independent testing across time, reducing dependence on spatially auto-correlated samples. The reference land-use dataset for 2020 was derived from supervised classification of Sentinel-2 imagery, with an overall accuracy exceeding 90% based on 300 manually interpreted ground-truth points. All inputs were harmonized to a 30 m resolution and spatially aligned for pixel-wise comparison.

The simulation results demonstrated strong alignment with observed patterns. The model achieved an Overall Accuracy (OA) of 91.3%, a Kappa coefficient of 0.82, and a Figure of Merit (FoM) of 0.64. The FoM—focusing on correctly predicted change cells—substantially outperforms the 0.35–0.55 range typically reported in conventional CA models, indicating high sensitivity to actual urban expansion zones (

Table 9. Model Accuracy Assessment (2010–2020 vs. 2010–2024).

Metric	2010–2020	2010–2024
Overall Accuracy (OA)	91.30%	89.70%
Kappa Coefficient	0.82	0.79
Figure of Merit (FoM)	0.64	0.60

). These results reflect the added explanatory power of functional-behavioral indicators derived from POI data.

Spatial overlay analysis confirmed that the model effectively captured key directions of urban expansion, especially southward growth near residential and institutional POI clusters and commercial densification along the riverfront. Overestimations were primarily found in the western corridor and peri-urban edges, possibly due to topographical misclassification or unrecorded speculative development.

To assess behavioral fidelity,

Table 10. Distribution of Simulated Probabilities for Real Transitions (2010–2020).

Probability Range	% of Real Transition Cells
>0.9	23.2%
0.8–0.9	26.4%
0.6–0.8	32.5%
0.4–0.6	12.3%
<0.4	5.6%

presents the distribution of transition probabilities for cells that experienced actual land-use change. Over 82% of these cells had predicted probabilities above 0.6, with nearly 50% exceeding 0.8. This suggests the behavioral logic of the model aligns well with real-world dynamics.

Table 11. Spatial Distribution of Overestimated Growth Predictions.

Zone Type	% of False Positives
Western Hilly Corridor	12.3%
Peri-urban Fringe	7.6%
Institutional Reserve Area	5.2%
Agricultural Flatlands	9.9%

summarizes the spatial distribution of false positives, highlighting error concentrations in the western hills and peri-urban fringe areas. This helps clarify where the model tends to overestimate urban expansion and reflects potential mismatches between functional drivers and topographic constraints.

These spatial concentrations of error reveal areas where functional signals may override topographic constraints or speculative development may not yet be registered in official data. Although a visual error map is not provided due to space and data size limitations, the structured summary in Table 11 enables effective spatial interpretation and identification of model bias zones. Comparative assessment between POI-enhanced and baseline models highlights the added value of POI data. As shown in

Table 12. Performance Comparison: POI-Enhanced vs. Baseline Model.

Indicator	With POI	Without POI
Kappa	0.82	0.68
FoM	0.64	0.47
OA	91.30%	86.50%

, removing POI variables significantly reduced performance.

The reliability of results across heterogeneous subregions—ranging from government administrative zones to emerging educational areas—demonstrates the robustness of the POI-informed framework. Nevertheless, spatial autocorrelation may still influence accuracy metrics. Although temporal separation of training and validation sets reduces this risk, future work could explore spatial k-fold cross-validation or residual-based spatial diagnostics to further enhance interpretability.

As described in Section 3.3, POI-based indicators—including density (PD), diversity (PDI), and functional centrality (FC)—were min–max normalized and embedded into the logistic regression model as continuous predictors. This design facilitates a behaviorally grounded simulation of spatial structure evolution.

Overall, the validation confirms that the proposed POI-enhanced CA framework is both technically reliable and behaviorally interpretable. It provides a robust basis for policy-sensitive scenario analysis, which is further explored in subsequent sections.

5.2. Spatial Patterns of Urban Expansion

The POI-enhanced CA simulation of Yan’an’s urban expansion from 2010 to 2024 reveals a distinctly corridor-oriented, polycentric, and functionally driven spatial evolution, diverging from the conventional model of concentric urban growth. The observed patterns reflect the dynamic interplay between topographic constraints, policy-guided infrastructure investment, and behaviorally encoded functional clustering.

As shown in Figure 6, the most intensive expansion occurred in the southern and southeastern quadrants of the study area. These zones, proximate to dense POI agglomerations and recently upgraded arterial roads, became focal points for new residential, educational, and commercial development. Over 40% of total urban land conversion during the simulation period occurred in these subregions, underscoring the alignment between functional demand, transport accessibility, and growth trajectories. Building upon the high simulation accuracy demonstrated in Section 5.1, these spatial manifestations offer not only geographic validation but also behavioral resonance with functional and infrastructural realities.

In contrast, the northern and northwestern sectors witnessed negligible growth despite their proximity to the historical core. These areas are delimited by ecological redlines and steep slopes (>25°), both of which were effectively encoded in the model’s environmental exclusion layer. Simulated transition probabilities in these zones remained consistently below 0.2, confirming the model’s sensitivity to hard constraints embedded in regulatory planning and terrain morphology.

A defining spatial characteristic of the simulated expansion is the linear intensification along the east–west transportation spine—particularly adjacent to Yan’an Avenue and the regional expressway system. This growth pattern deviates from classical radial expansion and instead materializes as ribbon-like corridors and elongated mixed-use belts. The emergence of such morphology reflects a synergistic coupling between accessibility advantages and functional centrality, especially in commerce and logistics-related POIs.

The model also identifies the rise in secondary polycentric nodes in peripheral zones, notably in the western and southwestern edges. These micro-urbanization clusters, anchored by educational and healthcare POIs, represent behavioral gravity centers that catalyzed localized development beyond the primary growth corridors. These subregions, characterized by elevated POI entropy and localized accessibility, demonstrate the decentralization of functional demand, rather than mere land spillover from the core.

Table 13. Key Functional Subregions and Corresponding Growth Characteristics.

Subregion	Dominant POI Type	Expansion Pattern	Topographic Constraint	Policy Designation
Southern Core	Residential + Commercial	Dense, Infill-Oriented	Low	Urban Growth Zone
Southeastern Belt	Educational + Residential	Linear Corridor	Low	Policy-Supported Zone
Western Fringe	Educational	Polycentric Cluster	Moderate	Transition Zone
Northern Sector	Mixed, Sparse	Minimal Growth	High	Ecological Redline

summarizes the dominant subregions of growth and their associated functional, topographic, and policy characteristics.

Functional intensification within the historical and administrative core areas is also notable. Despite minimal availability of raw land, these zones exhibited vertical redevelopment and land-use transformation driven by service-oriented POI clusters. As illustrated in Figure 7, high values of POI entropy—used as a proxy for mixed-use intensity—were spatially aligned with predicted micro-transitions, indicating that redevelopment pressure was more strongly associated with functional diversity than with simple density.

Statistical analysis confirmed that POI clusters with Shannon entropy > 1.5 had urbanization probabilities nearly twice those of mono-functional areas. This reinforces theoretical propositions in urban morphology that mixed-use environments enhance land desirability and accelerate transformation. Interestingly, slope thresholds played a modulating role: while slopes above 25° acted as hard constraints, moderate slopes (10~20°) with high functional signals often underwent transformation—highlighting that behavioral demand can partially override environmental friction in the presence of strong functional stimuli.

To further verify the spatial logic of urban expansion, Local Moran’s I was applied to identify statistically significant hotspots of land conversion. Four high-growth clusters emerged, as shown in

Table 14. High-Growth Hotspots Identified by Local Moran’s I Analysis.

Cluster Zone	POI Type	Moran’s I Z-Score	Policy Signal
/Southern Corridor	Residential	+3.24	Urban Growth Target
Southeastern Belt	Education	+2.98	Planned Campus District
Government Core	Civic/Public	+3.10	Redevelopment Initiative

. These zones not only exhibited high spatial autocorrelation of urban transitions but also corresponded to government investment corridors and high composite POI scores, validating the model’s ability to reflect both policy-driven incentives and functionally induced agglomeration.

In sum, the spatial outcomes of the simulation depict a structured yet adaptive urban growth model. Expansion is neither uniform nor random, but selectively concentrated in accessibility-rich corridors, policy-enabled districts, and functionally diverse subcenters. The integration of POI-based behavioral indicators enables the CA model to transcend purely morphological simulations, producing semantically enriched spatial dynamics that resonate with Yan’an’s real-world development logic and institutional planning framework. Such semantic richness and localized sensitivity would be difficult to achieve through morphology-based CA models alone, underscoring the irreplaceable role of POI-derived indicators in simulating heterogeneous urban expansion.

5.3. Functional Drivers Interpreted from POI Influence

The simulation results from 2010 to 2024 provide compelling evidence for the functional logic embedded within the POI-enhanced Cellular Automata model. Rather than treating POIs as peripheral indicators, the model demonstrates that POI-based metrics—particularly density, diversity, and centrality—serve as core explanatory variables in determining land-use transformation probability. The spatial distribution of new urban land closely aligns with regions exhibiting high functional pressure and complexity, reinforcing the proposition that urban expansion is not merely constrained by accessibility or physical suitability, but is actively driven by socio-spatial demand.

Among the three core POI indicators, POI Density (PD) exhibited the strongest primary correlation with land conversion. Grid cells in the top quartile of PD values were 1.8 times more likely to transition to built-up land than those in the bottom quartile, even after controlling for slope and proximity to infrastructure. This suggests that the sheer intensity of localized urban functions—commercial outlets, residential clusters, educational facilities—acts as a gravitational field for development. The impact of PD was particularly evident in peripheral sub-centers, where isolated but dense POI clusters catalyzed rapid infill and new development (

Table 15. POI Density Quartiles and Their Effect on Urban Land Conversion.

POI Density Quartile	Avg. Transition Probability	Relative Risk (vs. Q1)
Q1 (Lowest)	0.22	1.00
Q2	0.34	1.55
Q3	0.41	1.86
Q4 (Highest)	0.51	2.32

).

However, density alone did not account for all spatial variations. POI Diversity (PDI), measured using a Shannon entropy index, emerged as an independent and reinforcing influence. Cells with high PDI (entropy > 1.5) experienced faster and more sustained urban growth than those dominated by a single POI type. These zones often evolved into mixed-use enclaves, accommodating housing, commerce, education, and healthcare within compact footprints. The results indicate that functional complexity creates spatial synergies, reducing development friction and supporting efficient land turnover.

Further disaggregation by POI category confirmed differentiated impacts on spatial transformation. Commercial POIs showed the most potent local effects, particularly when co-located with transit infrastructure, acting as accelerators of conversion due to high consumer activity and investment momentum. Educational POIs, while less spatially intense, induced broader transformation zones—especially in newly developed southern peripheries where campuses, dormitories, and supporting retail emerged simultaneously. Civic and healthcare POIs contributed to growth stabilization rather than acceleration, shaping medium-density, policy-aligned developments.

The concept of Functional Centrality (FC)—derived from kernel density estimations of high-level POIs such as government buildings and hospitals—proved to be the strongest predictor of polycentric development. FC hotspots corresponded with the emergence of new urban nodes, especially in the western and northeastern regions of Yan’an. These zones were not mere extensions of the urban core but exhibited autonomous spatial momentum, absorbing expansion pressure and reshaping local hierarchies (

Table 16. Differential Influence of POI Functional Categories on Spatial Transformation.

POI Type	Average Transition Probability	Spatial Influence Pattern
Commercial	0.58	High-intensity, localized
Educational	0.51	Broad zone, sub-center forming
Residential	0.44	Transitional belt, infill
Medical	0.38	Stabilizing, moderate intensity
Governmental	0.41	Strategic, polycentric potential

).

Non-linear behavior was also observed. Cells with extremely high POI density did not always convert at higher rates. Saturation effects, infrastructure bottlenecks, and regulatory constraints appeared to dampen further transformation. Similarly, while diversity generally enhanced conversion, excessively high entropy was occasionally associated with reduced change, likely due to land scarcity or zoning restrictions.

Temporal dynamics further validated the interpretive strength of POI indicators. The model’s use of 2023 POI data to simulate transformations from 2015 to 2024 suggests that POIs can act as early-warning signals or proxies for latent urbanization forces. Zones where new educational or retail functions emerged in 2016–2018 often saw urban growth prior to official rezoning or infrastructural investment, emphasizing the anticipatory value of functional monitoring.

Another salient finding was the synergistic effect of co-located POIs. Zones where educational, commercial, and residential POIs overlapped showed transition probabilities 20–30% higher than single-function areas. These zones, behaviorally complete and spatially compact, formed the backbone of the simulation’s highest-growth areas. Such patterns support contemporary planning literature advocating for mixed-use intensity as a foundation for sustainable urban morphology.

Lastly, interactions between POI pressure and environmental controls highlighted the tension between behavioral drivers and spatial policy. While redlines and slope masks constrained most high-risk areas, pockets of moderate-slope zones with strong POI influence did witness conversion. This suggests that functional pressure can, under certain conditions, override moderate environmental resistance—a critical insight for integrated land-use policy design.

In conclusion, the integration of POI-derived indicators into the CA framework has significantly enriched the model’s interpretive depth and behavioral realism. By shifting the simulation paradigm from purely physical constraints to demand-driven functional dynamics, the model unveils not only where urban transformation occurs, but why and under what behavioral triggers. These insights provide urban planners with a powerful diagnostic lens to understand latent urbanization forces, identify emerging sub-centers, and preemptively align infrastructure provision with evolving socio-spatial demands. The POI metrics—density, diversity, and functional centrality—thus transcend their role as passive spatial correlates and become active indicators of urban behavioral structure, offering novel pathways for responsive and anticipatory spatial governance. These behavioral insights are not only theoretically meaningful but also directly applicable to early-warning systems, spatial zoning updates, and anticipatory infrastructure investment.

5.4. Scenario Simulation and Planning Implications

To assess the strategic robustness of the POI-enhanced Cellular Automata (CA) model under divergent spatial governance regimes, two contrasting simulation scenarios were developed: a Business-as-Usual (BAU) scenario and an Ecological Constraint (EC) scenario. These simulate urban growth from 2020 to 2024 based on differing assumptions—BAU extrapolates purely from functional demand, while EC incorporates formal spatial controls drawn from land-use plans, ecological redlines, and slope restrictions. This comparative framework allows for critical evaluation of how behavioral momentum and regulatory spatial containment interact to shape urban morphology and land-use efficiency.

The BAU scenario, unconstrained by zoning limits, projected a 21.4% increase in urban built-up area over the simulation period. Expansion predominantly occurred along southern and western corridors—zones of high POI density and emerging sub-centers—forming ribbon-like, dispersed morphologies. Development clustered around transit infrastructure, educational institutions, and residential-commercial POI conglomerates, reflecting the bottom-up pull of functional saturation.

Conversely, the EC scenario applied strict exclusion masks to ecological redlines, high-slope zones (>25°), and legally designated no-development parcels. The result was a reduced urban growth rate of 17.2%, but one that exhibited more spatial coherence and infill intensification. Growth in this scenario was redirected toward underutilized inner parcels and transit-accessible belts, yielding a more compact and efficient urban form (

Table 17. Comparative Performance Metrics: BAU vs. EC Scenarios.

Metric	BAU Scenario	EC Scenario
Urban Expansion Rate (2020–2024)	21.4%	17.2%
Avg. POI Density in New Urban Land	0.68 POIs/km2	0.79 POIs/km2
Avg. POI Entropy Index	1.27	1.33
% of Development in Infill Areas	41.6%	56.2%

).

Figure 8 illustrates the divergent spatial morphologies under both scenarios. The BAU output features edge-expansion and corridor elongation, whereas the EC outcome reveals node-based intensification and polycentric densification.

While the EC scenario restricted outward growth, it did not fully suppress latent development pressure. Functional hotspots beyond the urban containment boundary—especially near emerging retail and service clusters—still exhibited elevated transition probabilities. These zones, despite regulatory exclusion, represent behaviorally desirable growth frontiers. If ignored in planning, such areas are prone to informal or unauthorized development.

This spatial misalignment between policy and behavior reveals a critical governance challenge. Regulatory regimes that fail to account for functional intensity—proxied here by POI density and entropy—may inadvertently displace urban pressure into ecologically fragile or infrastructurally unprepared zones. The model suggests that where behavioral attractors exist, purely spatial prohibitions may prove insufficient.

Moreover, the EC scenario demonstrated a functional-efficiency trade-off. Despite reduced spatial expansion, it achieved a 16.7% higher average POI density in newly developed cells and a significant increase in mixed-use share. These indicators affirm that compactness, if behaviorally guided, can enhance service integration and spatial performance (

Table 18. Functional Efficiency Indicators under Each Scenario.

Indicator	BAU	EC
Average Distance to Major POIs	1.42 km	1.09 km
Share of Mixed-Use Grid Cells	27.8%	39.5%
Functional Co-location Score	0.38	0.51

).

These findings yield several critical implications for urban spatial governance.

First, functional demand is spatially resilient—it persists and intensifies even in zones formally excluded from development. Regulatory barriers, such as ecological redlines or slope-based exclusions, may delay expansion but do not eliminate latent growth pressure. If such areas—particularly those with high POI density and functional entropy—are ignored, they risk becoming hotspots for informal growth, undermining both ecological protections and planning legitimacy. This necessitates a transition from rigid exclusion to adaptive redirection, employing tools such as strategic upzoning, Transit-Oriented Development (TOD), or land readjustment mechanisms to channel demand without eroding regulatory intent.

Second, the simulation confirms the value of functional mapping as a preemptive governance instrument. By translating POI-derived indicators into probabilistic forecasts of land conversion, planners can identify incipient development zones before morphological changes occur. This enables anticipatory infrastructure deployment, proactive policy design, and spatial interventions that align institutional timelines with behavioral dynamics—narrowing the chronic lag between “where people go” and “where policy permits.”

Third, the EC scenario demonstrates that ecological constraint and functional efficiency are not mutually exclusive. Compact, infill-oriented development—guided by functional clustering—can reinforce ecological protection by minimizing sprawl and infrastructure costs. The model, thus, supports a governance shift from binary growth control (permit vs. prohibit) to a more nuanced paradigm of “functionally guided containment”, where land-use decisions are informed by both ecological thresholds and urban behavioral logic.

Finally, the POI-CA framework proves useful not only as a simulation engine but as a behaviorally anchored decision-support system. It enables planners to visualize trade-offs, stress-test zoning levers, and evaluate how development might unfold under diverse policy regimes. For mid-sized, topographically constrained cities such as Yan’an, this integrated approach offers a roadmap toward reconciling development ambition with spatial and ecological responsibility—by building governance not against behavior, but through it.

6. Discussion

6.1. Behavioral Mechanisms and Functional Sensitivities in Urban Simulation

The integration of POI-based indicators into the Cellular Automata framework offers a new lens through which to interpret the behavioral logic of urban spatial expansion. Unlike conventional simulations that rely primarily on accessibility, slope, or historical trend extrapolation, the POI-enhanced model captures the functional attributes of urbanization—density, diversity, and centrality—reflecting not only where development occurs but why it concentrates in specific locations.

The simulation results between 2010 and 2024 reveal that urban growth in Yan’an was not randomly distributed, nor was it solely constrained by environmental or infrastructural boundaries. Rather, growth patterns followed clear functional gradients. Areas with high POI density—particularly those containing clusters of commercial, educational, and residential facilities—exhibited significantly higher probabilities of land conversion. This finding underscores the gravitational pull of functional intensity: as functional saturation increases, so does the spatial desirability and development momentum.

Among the POI indicators tested, POI Density (PD) exerted the most dominant influence. Grid cells in the top quartile of PD values were 2.3 times more likely to be transformed into urban land than those in the lowest quartile, even after controlling for slope and distance to roads. This suggests that localized service intensity—such as schools, markets, or healthcare—plays a primary role in shaping urban behavior. In peripheral sub-centers, isolated but dense POI clusters acted as magnets for infill growth, supporting the notion of emergent decentralization driven by socio-spatial demand rather than top-down planning.

In addition to density, POI Diversity (PDI)—measured through Shannon entropy—contributed an independent and reinforcing influence on transformation. Mixed-use environments with high entropy values (>1.5) exhibited stronger and more sustained development trajectories than functionally homogeneous zones. These areas often evolved into compact, multifunctional enclaves, where housing, commerce, and public services co-existed within walkable distances. The spatial synergies created in these zones effectively lowered development friction, improved land turnover, and optimized infrastructure use.

Further differentiation of POI types demonstrated heterogeneous behavioral effects. Commercial POIs showed strong local accelerative effects, especially when co-located with transport hubs. Educational POIs had a more diffuse impact, forming broader transformation zones—particularly in newly developing southern districts. Civic and medical POIs tended to stabilize rather than accelerate development, aligning with government policy on orderly growth and service provisioning. These differentiated impacts highlight the value of incorporating functional typologies into simulation logic, allowing for more granular interpretations of urban change.

One of the most significant conceptual contributions is the identification of Functional Centrality (FC) as a predictive factor for polycentric development. Derived from the kernel density of higher-order POIs—such as government offices and hospitals—FC hotspots closely corresponded to newly emerging nodes outside the historical urban core. These areas were not merely peripheral spillovers but demonstrated autonomous spatial momentum, characterized by clustering, service accumulation, and internal cohesion. As such, FC offers a spatial-behavioral measure of emerging urban hierarchy and node formation, complementing traditional centrality metrics based on transport or geography.

Non-linear dynamics also emerged. In highly saturated zones—particularly in central districts—the transition probability did not continue to rise with increasing POI density. Instead, a plateau or even decline was observed. This saturation effect likely results from regulatory caps, limited developable land, or infrastructural bottlenecks. Similarly, extremely high entropy values were occasionally associated with stagnant growth, possibly due to over-fragmentation of land use or zoning rigidities. These findings indicate that while functional intensity and diversity drive growth, they are bounded by systemic constraints and diminishing returns.

Temporal simulation further validated the behavioral explanatory power of POI indicators. The use of 2023 POI data to simulate land conversion from 2015 to 2024 revealed that emerging functions—such as new university campuses or retail corridors—predicted urban transformation even before formal rezoning or infrastructure rollout. In this sense, POIs serve as early-warning signals of latent urbanization, offering planners a behavioral foresight tool that complements static master plans.

Another notable finding lies in the synergistic effect of functional co-location. Grid cells containing overlapping POIs from at least three different categories (e.g., commercial, residential, educational) exhibited transition probabilities 20–30% higher than mono-functional cells. These areas—characterized by compactness, behavioral completeness, and service-rich environments—formed the backbone of high-growth zones in both baseline and scenario-based simulations. This supports broader urban planning literature advocating for mixed-use intensity as a foundational element of sustainable urban morphology.

Finally, the interaction between functional drivers and ecological constraints revealed critical governance tensions. In the Ecological Constraint (EC) scenario, while high-slope and redline zones were largely protected, certain medium-risk areas with strong functional signals did undergo conversion. This suggests that functional gravity can, under certain conditions, override moderate environmental resistance—a critical insight for planning in rapidly urbanizing but ecologically fragile contexts like Yan’an. It underscores the importance of aligning ecological protection zones with functional demand maps to avoid enforcement dilemmas or informal development.

In summary, the behavioral mechanisms extracted from POI-based simulations provide a multi-dimensional interpretation of urban growth. They validate the hypothesis that urban expansion is not simply a function of geography or policy, but a spatial manifestation of latent socio-economic demand embedded in functional configurations. The integration of these indicators into the CA framework enhances not only model accuracy but also interpretability, offering a more behaviorally responsive and practically relevant tool for planners. This fusion of data-driven behavioral insight and spatial modeling marks a significant advancement in urban simulation, with direct implications for predictive planning, zoning reform, and urban governance.

6.2. Rethinking Spatial Governance: From Static Regulation to Functional Responsiveness

The integration of POI-based functional indicators into simulation frameworks does more than improve prediction accuracy—it fundamentally challenges prevailing assumptions about how spatial governance should be structured, executed, and evaluated. The contrasting outcomes observed in the Business-as-Usual (BAU) and Ecological Constraint (EC) scenarios underscore a deeper issue: the persistent mismatch between regulatory spatial frameworks and functional urban dynamics.

Traditional urban governance relies heavily on static spatial instruments—such as zoning maps, redline boundaries, and growth containment lines—to exert control over urban form. These instruments assume that planning agencies can accurately forecast future demand and that urban actors will conform to administratively imposed spatial logics. However, the behavioral tendencies revealed by the POI-enhanced CA model expose the limitations of this regulatory paradigm. Urban expansion in Yan’an did not halt at the edge of a boundary line; rather, it responded to accumulations of service functions, activity density, and accessibility—factors often decoupled from policy maps.

One critical insight from the EC scenario is that even functionally intense zones beyond growth boundaries continue to exhibit strong transition probabilities. These “behavioral hotspots” signal active human demand, but without regulatory recognition, they risk becoming sites of informal development or governance failure. This misalignment reveals that current governance models, which prioritize spatial exclusion over behavioral integration, are increasingly inadequate in fast-changing urban environments.

The results advocate for a shift from static, exclusion-based governance toward an adaptive, demand-sensitive model. This new model requires that land-use planning incorporate real-time or near-real-time indicators of functional pressure—such as POI entropy, density gradients, or co-location scores—into both short-term decisions (e.g., land allocation, zoning variance) and long-term strategies (e.g., infrastructure investment, greenbelt redesign).

Such a paradigm shift aligns with the concept of “responsive spatial governance”, wherein regulatory frameworks are continuously updated based on evolving socio-functional configurations. Rather than enforcing rigid compliance with outdated plans, responsive governance seeks to anticipate latent urbanization and guide it toward sustainable and infrastructure-ready locations. This anticipatory approach is especially critical in mid-sized cities like Yan’an, where growth is not only rapid but spatially fragmented, and where governance capacity may be constrained.

Another governance implication concerns the evaluation of planning success. Conventional performance metrics—such as containment effectiveness or growth rate control—fail to capture the qualitative outcomes of planning decisions. By contrast, the POI-based functional efficiency metrics presented in this study (e.g., average distance to major POIs, mixed-use cell ratio, functional co-location score) offer a richer, multidimensional framework for assessing the spatial quality and social responsiveness of urban growth.

For example, under the EC scenario, although total expansion area was reduced, the efficiency and coherence of land-use significantly improved. This suggests that regulatory control, when informed by functional signals, can not only limit sprawl but also enhance spatial productivity. Therefore, instead of asking “how much land was saved,” governance evaluation should also ask “how well was development matched to function and need.”

The findings also point to a rethinking of growth boundaries. Current boundary-setting practices are typically based on ecological or physical constraints (e.g., topography, hydrology) and population projections. While valid, these criteria often ignore emergent socio-spatial logics. The POI-enhanced simulation reveals that behavioral demand precedes regulatory recognition: POI accumulation patterns anticipate urban transformation, not the other way around. This suggests that future growth boundaries should be more elastic and data-informed, using POI and other activity indicators as early-warning systems for potential expansion zones.

Importantly, governance reform must consider the equity dimension. When high-functioning zones outside boundaries are excluded from development, they may become targets for informal expansion—unregulated, underserved, and often environmentally risky. This generates a “functional injustice,” wherein behaviorally driven demand is systematically ignored, leading to spatial inequality in access to infrastructure, transport, and public services. Responsive governance must proactively integrate these zones into planning frameworks, ensuring that formal recognition accompanies functional growth.

In institutional terms, these findings demand closer coordination between behavioral analytics and spatial regulation. Planning departments must move beyond static GIS overlays and embrace dynamic functional datasets—POI, mobile signaling, smart city sensors—as core inputs in governance decisions. Moreover, regulatory tools (e.g., zoning laws, land use permits, infrastructure budgets) must be recalibrated to reflect not just spatial form but functional vitality.

This also implies a need for capacity building within local planning agencies. Data literacy, simulation modeling, and scenario-based decision-making must become routine capabilities. For many second-tier cities like Yan’an, this requires institutional innovation and the establishment of cross-departmental planning units capable of translating behavioral data into regulatory action.

Finally, the results speak to the broader epistemological transformation of planning: from a regulatory art rooted in prescriptive control to an interpretive science grounded in behavioral pattern recognition. The POI-CA model acts not merely as a technical simulation tool, but as a governance diagnostics platform, enabling planners to visualize where, why, and how spatial tension arises—and how it might be resolved through targeted intervention.

In summary, the integration of behavioral dynamics into spatial governance transforms how we define planning problems, anticipate risk, and measure success. For cities facing the dual pressures of ecological protection and urban expansion, responsive, data-informed, and functionally aligned governance models offer a pathway toward sustainable spatial futures. Rather than resisting behavioral logic, urban governance must evolve to engage, redirect, and optimize it—aligning institutional goals with human geography.

6.3. Planning Implications and Theoretical Contributions

This study yields several planning implications and theoretical contributions that extend beyond conventional CA-based urban simulation and enrich the discourse on behaviorally informed spatial governance.

First, from a practical planning perspective, the results emphasize the necessity of integrating behavioral-functional indicators into land-use regulation. The POI-enhanced CA model reveals that urban expansion is not only a product of top-down zoning and infrastructure provision but also a bottom-up response to functional intensity and socio-spatial demand. Planning systems that fail to account for such latent forces risk policy blind spots, leading to informal growth, inefficient land conversion, or governance friction. By mapping POI-based indicators such as entropy, density, and centrality, planners can pre-emptively identify zones of emergent pressure and revise development rights, urban growth boundaries, or infrastructure plans accordingly.

Second, this study offers an operationalizable pathway for aligning ecological protection with urban growth. Contrary to the perceived trade-off between conservation and development, the EC scenario demonstrated that ecological constraints can be leveraged to promote functionally efficient urban forms. Growth, when steered toward areas of high functional saturation, can yield higher co-location efficiency, reduced travel distances, and denser mixed-use fabrics. This reframes spatial containment not as suppression, but as a tool of functional optimization—an approach particularly suitable for cities like Yan’an that are topographically and environmentally constrained.

Third, the findings propose a theoretical shift in how urban simulation models conceptualize transformation. Traditional CA frameworks often treat functional data as static environmental variables or peripheral constraints. In contrast, this study positions POIs as behavioral signatures—active agents that encapsulate demand, accessibility, investment logic, and social utility. By embedding these indicators into the CA mechanism, the model not only improves predictive accuracy but also renders urban change more interpretable and actionable. This contributes to the emerging school of thought on behavioral urbanism, which emphasizes cognition, interaction, and functional use over merely geometric or administrative spatial logics.

Finally, this study contributes to planning theory by reinforcing the value of adaptive governance models. The divergence between BAU and EC scenarios illustrates that rigid spatial regulation often underestimates the flexibility and momentum of urban systems. Planning frameworks must, therefore, transition from static zoning regimes to anticipatory systems that incorporate real-time data, dynamic simulation, and feedback-informed revision cycles. The POI-CA framework proposed here can serve as a foundational tool in such adaptive cycles, allowing planners to test policy outcomes, simulate alternative trajectories, and fine-tune interventions across multiple governance scales.

In sum, the integration of functional-behavioral logic into spatial simulation opens a new frontier in both theory and practice. It bridges the gap between socio-economic demand and spatial regulation, offering a decision-support system grounded not only in physical constraints but also in the evolving logic of urban use. This paradigm holds promise for more responsive, sustainable, and inclusive urban governance in rapidly transforming mid-sized cities worldwide.

7. Conclusions

This study proposes a behaviorally enriched urban simulation framework that integrates functional POI indicators with a probabilistic Cellular Automata (CA) model to analyze the spatio-temporal evolution of land-use structure in Yan’an from 2010 to 2024. By combining high-resolution land-use data, semantically processed POI metrics, and scenario-based simulation, the model captures not only where urban expansion occurs, but also why specific locations undergo transformation—providing both spatial precision and behavioral interpretability.

Empirically, the model achieved high spatial fidelity, with an Overall Accuracy (OA) of 91.3%, a Kappa coefficient of 0.82, and a Figure of Merit (FoM) of 0.64, exceeding performance benchmarks of traditional CA models. The integration of POI density, entropy, and accessibility variables significantly improved simulation performance. It allowed the model to replicate observed growth patterns—such as southward corridor expansion, clustering around educational and commercial hubs, and avoidance of topographically or ecologically constrained zones—with greater explanatory logic. Importantly, the inclusion of POI-based functional signals revealed the underlying spatial logic of transformation, capturing the latent forces behind urban land-use change.

Functionally, the model illustrates how land-use conversion is strongly shaped by behavioral attractors—particularly high POI intensity and mixed-function configurations. It demonstrates that functional centrality can foster polycentric growth beyond the historical urban core, while socio-functional momentum may override moderate physical constraints such as slope or distance to infrastructure. These findings reposition urban expansion not as a purely morphological outcome but as the emergent result of layered behavioral demand and institutional opportunity structures.

The dual-scenario simulations—Business-as-Usual (BAU) and Ecological Constraint (EC)—further reveal the tensions and trade-offs between functional demand and regulatory containment. While the BAU scenario highlights the risk of fragmented, infrastructure-misaligned growth when development follows functional inertia alone, the EC scenario illustrates how ecological policies, if aligned with latent functional structure, can channel expansion toward compact, efficient, and sustainable urban forms. These results call for a more adaptive governance paradigm—one that responds to real-time functional signals rather than static zoning templates.

Theoretically, this research contributes to a new generation of urban simulation that transcends morphology-based modeling by embedding semantic, behavioral dimensions. POI indicators—often dismissed as auxiliary—are elevated here to core proxies of urban vitality and spatial desirability. This repositions urban land-use change as a functionally contingent, behaviorally interpretable, and governance-malleable process. The integrated modeling architecture—linking logistic regression, CA transition rules, and POI semantics—offers a transferable, scalable method for other fast-growing or topographically constrained urban contexts.

Practically, the study provides planners with a decision-support tool capable of revealing emergent development pressures, optimizing spatial efficiency, and aligning ecological protection with socio-economic responsiveness. It demonstrates that real-time POI monitoring can serve as a proxy for anticipatory planning and that planning interventions should incorporate functional pressure maps to preempt informal sprawl or regulatory mismatches.

Nonetheless, several methodological limitations warrant further discussion. First, although the integration of POI data introduces behavioral granularity into spatial simulation, the current implementation relies on static POI snapshots, which may not fully capture the temporal dynamics of urban functions. Urban vitality is inherently fluid, and future research could incorporate time-series POI datasets or real-time data streams (e.g., map queries, foot traffic) to better model the co-evolution of function and form.

Second, the CA model operates in two-dimensional space and simulates land-use transitions at the surface level. In mountainous cities like Yan’an, vertical expansion (e.g., floor area, building heights) constitutes a significant component of urban growth. Extending the model to incorporate 3D built environment data, such as from LiDAR or building footprints, would greatly enhance realism and planning relevance.

Third, while our POI-based CA approach effectively simulates aggregate patterns and functional trends, it abstracts from agent-level heterogeneity and institutional feedback. Alternative approaches, such as agent-based models (ABM) or multi-agent hybrid frameworks, could better represent decentralized decision-making, negotiation processes, and stakeholder-specific behaviors that shape urban form.

Additionally, the model does not compare directly with other popular machine learning algorithms such as Random Forest (RF), Support Vector Machines (SVM), or Gradient Boosted Decision Trees (GBDT). These methods have demonstrated high predictive accuracy in land-use classification and change detection. However, unlike these “black-box” classifiers, the POI-CA framework retains spatial transparency, interpretability, and policy-actionability—qualities that are essential for urban planning scenarios. That said, future studies could conduct side-by-side comparisons or ensemble integrations to better assess trade-offs between accuracy and interpretability.

Finally, the model assumes stable relationships between POI indicators and land-use conversion likelihood. Yet in reality, such relationships may evolve under new infrastructure, institutional shocks, or policy shifts. To address this, future work could adopt dynamic transition functions, feedback loops, or deep learning architectures that update behavioral weights over time.

In sum, this study offers a behaviorally informed and policy-relevant simulation architecture that can guide adaptive planning. Yet acknowledging its limitations also points toward a richer agenda of technical, theoretical, and participatory enhancements for future spatial modeling research.