GIS-Based Analysis of Retail Spatial Distribution and Driving Mechanisms in a Resource-Based Transition City: Evidence from POI Data in Taiyuan, China

Abstract

Rapid urbanization in China has reshaped retail spatial structures, creating challenges of accessibility and service equity. This study employs a Geographic Information Systems (GIS)-based analytical framework to examine the spatial distribution and driving mechanisms of retail outlets in Taiyuan, a resource-based transition city in central China. Using 2023 Point of Interest (POI) data and a 2 km × 2 km grid system, kernel density estimation (KDE), Average Nearest Neighbor (ANN) Analysis, Location Quotient (LQ), and spatial autocorrelation were applied to identify clustering patterns and functional specialization. The GeoDetector (Word version, downloaded 2025) model further quantified the explanatory power of twelve natural, social, economic, and transportation variables. Results reveal a polycentric retail structure, with high-density clusters in Yingze and Xiaodian districts and under-supply in Jiancaoping and Jinyuan. Population density, nighttime light (NTL) intensity, and school distribution emerged as the strongest drivers, while topography constrained expansion. By integrating GIS-based spatial statistics with GeoDetector, the study demonstrates a transferable framework for analyzing urban retail spatial patterns. The findings extend retail geography to transition cities and provide practical guidance for optimizing retail allocation, enhancing service equity, and supporting spatial decision-making for sustainable urban development.

1. Introduction

With the rapid development of geospatial multi-source spatial data technologies such as POI and remote sensing, urban spatial economic research can now capture fine-scale commercial dynamics and spatial interactions with greater precision, enabling refined analyses of urban retail systems [1,2,3,4]. Currently, Chinese cities have entered a new stage of spatial restructuring and endogenous growth, and the spatial pattern of commercial space has evolved from single-center agglomeration to multi-center networking [5,6], reflecting the diversification of retail centers and functional nodes in the urban core. As a micro-representation of urban spatial structure, the distribution of retail outlets is not only related to the convenience and fairness of residents’ access to services but also reflects the efficiency of urban resource allocation and spatial justice [7,8].

In recent years, the integration of point-of-interest (POI) data and geographic information systems (GIS) has greatly improved the precision and breadth of analyzing the spatial structure of urban commerce [9,10]. However, most of the current research focuses on developed cities in the east, such as Beijing, Shanghai, and Guangzhou, and the research topics are mainly centered on the central place theory [11], the hierarchical structure of commercial business, commercial agglomeration, and the ‘15-min living circle’ [12,13]. In contrast, empirical discussions on the retail spatial structure and distribution mechanism of resource-based transition cities in central and western China are still relatively weak [14,15].

Resource-oriented cities are urban areas historically dominated by energy or mineral industries that are currently undergoing economic transformation and industrial diversification. Such cities face challenges in converting industrial land to commercial functions and restructuring employment structures. Taiyuan exemplifies this transformation as it transitions from a coal-dependent economy to a service-oriented metropolitan hub [16,17].

As the capital of Shanxi Province and a typical resource-oriented city, Taiyuan is at a critical stage of transformation from an industry-oriented to a service-oriented economic structure. Its main urban areas (Yingze, Xiaodian, Wanbailin, Jiancaoping, Xinghualing and Jinyuan districts) have experienced significant urban renewal and commercial function restructuring in recent years, exhibiting a spatial distribution trend from clustering along main roads to diversified area penetration. In this context, exploring how retail outlets can reconfigure their spatial distribution pattern under the combined effects of topographical constraints, transportation accessibility and population changes has become an important issue to be solved.

This study takes Taiyuan City as an example, focusing on the spatial distribution and agglomeration characteristics of its retail outlets and exploring the multiple driving mechanisms behind them. The full sample data of points of interest (POI) for 2023 were obtained from the Amap (Gaode) Open Platform API, which provides comprehensive geographic coordinates and categorical information for retail outlets. Based on these data, a 2 km × 2 km spatial grid was constructed to analyze the degree of agglomeration of various retail categories (such as restaurants, home appliances, and general supermarkets) and their spatial differentiation [7,13,18]. Kernel density analysis, Average Nearest Neighbor (ANN) Analysis, location quotient, and the GeoDetector model were jointly applied to reveal spatial clustering characteristics and influencing factors. The GeoDetector (Word version, downloaded 2025) model is a spatial statistical tool that quantifies how explanatory factors influence the spatial distribution of a dependent variable through the q-statistic [19]. Meanwhile, natural and socioeconomic variables such as population density, GDP level, topographic relief, and transportation facilities were introduced to assess their explanatory power and interactive effects on retail distribution [18,20].

This study has three academic contributions: first, from the perspective of micro-spatial units, it systematically reveals the differentiated agglomeration patterns presented by different types of retail outlets in urban space, which enriches the spatial typology study of the urban commercial system; second, it combines the four categories of factors, namely, natural, social, economic, and transportation, to construct a causal explanatory framework for the spatial distribution of retail outlets, which enhances the understanding of the mechanism of spatial differentiation of the city [21,22]; third, based on spatial modeling and visualization analysis, it provides actionable policy recommendations for the optimization of commercial spatial structure in resource-based cities, helping to achieve the goals of urban equity and sustainability [23,24].

Overall, this study focuses on a city type that has received relatively little attention in current urban geography research, which not only expands the body of empirical research on retail spatial patterns in central cities, but also provides theoretical and practical support for business planning and spatial governance in cities in economic transition.

2. Literature Review and Theoretical Framework

Research on the spatial distribution of urban retail outlets has evolved from classical location theories to data-driven analyses leveraging geospatial big data. Early studies, such as Christaller’s central place theory and Lösch’s spatial economic model, provided foundational insights into hierarchical and cost-demand-based spatial organization of commercial facilities [11,25]. With urban systems shifting from monocentric to polycentric forms, subsequent studies emphasized the agglomeration mechanisms of economic activities in complex urban networks [3,15]. However, these frameworks are often insufficient to fully capture the dynamic, multi-nodal retail patterns in contemporary cities, particularly in resource-transition regions [2,26,27].

The integration of point-of-interest (POI) data with Geographic Information System (GIS) techniques has substantially enhanced the precision and granularity of retail spatial analysis. POI-based approaches allow detailed identification of outlet types, density patterns, and clustering tendencies, supporting meso-scale assessments of urban commercial structure [9,10,12,28]. Analytical methods such as kernel density estimation (KDE), Average Nearest Neighbor (ANN) Analysis, and Location Quotient (LQ) are widely employed to quantify retail agglomeration, functional specialization, and spatial heterogeneity [22,29,30].

Recent studies also highlight the multidimensional factors influencing retail spatial distribution. Natural conditions, including elevation and slope, constrain outlet locations in hilly cities [7,21]. Social variables, such as population density, schools, and hospitals, shape local demand and attract retail clustering [23,31]. Economic indicators, including GDP, housing prices, and company and enterprise density, influence the location of high-end retail outlets [29,30,31]. Transportation accessibility, encompassing road networks and transit hubs, further correlate with consumer reach and commercial agglomeration [7,12,31,32,33]. Importantly, the interaction of these factors rather than their isolated effects often drives the observed spatial patterns, which can be effectively assessed using statistical tools such as the GeoDetector (Word version, downloaded 2025) model [8,13,18,34,35].

The theoretical foundation of this study draws on central place theory, the first law of geography, and new economic geography concepts, providing complementary perspectives on hierarchical organization, spatial proximity, and scale economies in retail distribution [11,26,27]. Combining these perspectives with GIS-based analytical methods enables a systematic examination of both the spatial patterns and driving mechanisms of retail outlets in Taiyuan, particularly under the constraints of urban transformation, terrain complexity, and the spatial distribution of population [33,34,35,36,37].

In summary, the literature suggests three key insights guiding this study: first, retail spatial patterns are increasingly polycentric and heterogeneous, reflecting complex urban dynamics; second, POI and GIS methodologies offer precise, multi-scale tools including kernel density estimation (KDE), Average Nearest Neighbor (ANN) Analysis, LISA, and GeoDetector, that enable pattern recognition and factor detection across multiple spatial levels; and third, multidimensional and interacting factors—including natural, social, economic, and transportation elements—shape retail spatial distributions, which can be effectively quantified using models such as GeoDetector (Word version, downloaded 2025). These insights form the basis for the subsequent empirical analysis of Taiyuan’s retail landscape [31,34,38].

However, despite these advances, two key research gaps remain. First, existing studies have predominantly focused on metropolitan or coastal cities, leaving resource-based transition cities like Taiyuan underexplored in terms of retail spatial mechanisms and policy implications [14]. Second, prior research often relies on single-scale or linear models, lacking a multidimensional and interactive framework to capture spatial heterogeneity and cross-factor dynamics [12,13]. Addressing these gaps, this study integrates POI-based spatial analytics and the GeoDetector (Word version, downloaded 2025) model to systematically quantify the multi-factor drivers shaping retail spatial distribution in Taiyuan, thereby extending the empirical scope and methodological depth of urban retail geography.

3. Research Methods and Data

3.1. Research Area and Research Object

This study focuses on the main urban area of Taiyuan, the capital of Shanxi Province, encompassing six administrative districts: Xiaodian, Yingze, Xinghualing, Wanbailin, Jiancaoping, and Jinyuan (Figure 1). Taiyuan, as a resource-oriented city, is undergoing a spatial transformation from a single-center agglomeration to a multi-center network, driven by the ‘transformation and comprehensive reform’ strategy [18,32,39]. This distribution is reflected in the rapid expansion of retail outlets and their increasingly diverse spatial and functional characteristics.

Geographically, Taiyuan is situated in the Fenhe Valley Basin with a ‘high west–low east’ terrain, influencing retail distribution [40]. Elevation restricts retail layouts in mountainous western and northern regions, resulting in lower outlet densities compared to the southeastern plains. The Fen River serves as both an ecological corridor and spatial boundary, shaping commercial clusters concentrated in Yingze and Xinghualing districts [39].

With ongoing urban expansion and transportation upgrades, Xiaodian and Jinyuan have emerged as new sub-centers, hosting major developments such as Changfeng Business District and South Railway Station Business District, contributing to a ‘multi-node–multi-axis’ commercial structure [8]. Kernel density and LISA were applied to identify cluster locations, while GeoDetector was employed to quantify and further confirm the extent to which each factor contributes to retail agglomeration [13].

The 46 street offices exhibit trends of old city renewal, suburban expansion, and industrial transformation. Streets such as Xuefu, Changfeng, and Longcheng have become spatial backbones for commercial development. Transportation accessibility and population distribution are key determinants of clustering [7,24,41].

Taiyuan’s 14th Five-Year Plan promotes a three-tier urban consumption system: municipal centers, regional districts, and community outlets. A 2 km × 2 km fishnet grid was selected after comparison with 1 km and 3 km alternatives. The 2 km resolution achieved the best balance between spatial detail and computational efficiency, consistent with previous studies on medium-sized Chinese cities [18,42].

To correlate with an appropriate analytical scale for intra-urban retail patterning, three grid sizes (1 km, 2 km, 3 km) were compared prior to formal analysis. The 1 km grid produced a high proportion of empty cells (approximately 35–40%), leading to unstable variance structures and fragmented KDE and LISA surfaces. In contrast, the 3 km grid substantially reduced spatial heterogeneity and overly smoothed the core–periphery gradient, masking meaningful sub-center variations.

The 2 km × 2 km grid achieved the best balance: it exhibited a low empty-cell rate (around 10%), captured major agglomeration centers without fragmentation, and maintained computational efficiency across clustering and spatial-autocorrelation procedures. A supplementary comparison showed that 2 km grids yielded the highest variance-explained ratio when regressing retail density on key socioeconomic drivers (population, NTL), further supporting scale appropriateness.

Although the Modifiable Areal Unit Problem (MAUP) cannot be completely eliminated, the observed spatial patterns were stable across neighboring scales, and 2 km provides an analytically robust and policy-relevant unit for describing intra-city retail structure. A brief comparison of grid-size performance is included in Appendix A 

Table A1. Comparison of Grid Sizes for Spatial Analysis.

Criterion	1 km Grid	2 km Grid (Final)	3 km Grid
Empty-cell rate	High (~35–40%)	Low (~10%)	Medium (~20–25%)
Detail level	Very high (fragmented)	Balanced	Low (over-smoothed)
Variance explained (density vs. drivers)	Low	Highest	Moderate
Computation time	Slow	Moderate	Fast
KDE/LISA clarity	Fragmented, noisy	Clear polycentric pattern	Excessively smooth
Suitability	Too fine	Optimal	Too coarse

.

The city’s urban transformation, complex terrain, and evolving functions make it an ideal case to study retail spatial patterns under multi-scalar constraints such as administrative boundaries, transportation corridors, and topographic barriers affecting commercial distribution. Focusing on these six districts allows a nuanced understanding of spatial distribution mechanisms and commercial layout.

3.2. Data Sources and Pre-Processing

All socioeconomic data, including population density, were obtained from the Taiyuan Statistical Yearbook2024 [43] which reports official data for 2023. Accordingly, all datasets used in this study correspond to the same reference year to ensure temporal consistency.

Retail POI Data: Obtained from Amap (Gaode) Map Open Platform (Version 2023), including outlet name, category, coordinates, and street-level attribution. Data were restructured according to the National Economy Industry Classification Standard (GB/T 4754-2017) [44], resulting in eight retail categories. Anomalies such as missing coordinates and unclassified entries were removed [9,28].

To ensure data reliability, all retail POIs (total 21,862 retail outlets) were cleaned and validated through duplicate removal, coordinate verification, and attribute checks. The detailed cleaning results are summarized in

Table 1. Data-cleaning process for retail POIs in Taiyuan.

Cleaning Step	Description	Number of Records Removed	Remaining POIs
Duplicate removal	Identical or near-identical name–address–coordinate combinations	512	21,350
Coordinate validation	Records with missing or invalid coordinates (e.g., outside boundary)	4	21,346
Attribute sanity checks	Records with incomplete or placeholder names	11	21,335
Category harmonization	Reclassified according to the National Economy Industry Classification Standard (GB/T 4754-2017)	-	21,335

.

After cleaning, a total of 21,335 valid retail POIs were retained and categorized into eight retail categories based on the GB/T 4754-2017 standard.

Socioeconomic Data: Derived from Taiyuan’s Statistical Yearbook, Shanxi Statistical Yearbook, and the Fifth National Economic Census. Variables include population density, company and enterprise density, nighttime light (NTL) intensity, and housing prices. Economic vitality was represented by nighttime light (NTL) data instead of GDP statistics, as the Taiyuan Statistical Yearbook provides GDP data only at the district scale, which is insufficient for intra-urban spatial analysis [45]. The NTL data derived from VIIRS 2023 composite imagery sourced from the National Oceanic and Atmospheric Administration (NOAA). This substitution ensures spatial continuity at the grid level for regression and GeoDetector analysis. All data layers were spatially aggregated and aligned to a uniform 2 km × 2 km grid using centroid-based spatial joins to ensure consistency across datasets.

Spatial Data: Road networks, transit stations, and parking lot densities were collected from OpenStreetMap and local bureaus. Topographical data (DEM, slope, ruggedness) were obtained from the ASTER Global Digital Elevation Model (GDEM) and matched to grids using ArcGIS Pro (Version 10.6) [18,46,47].

All datasets were projected to WGS_1984_UTM_Zone_49N, cleaned by removing duplicate POI entries and correcting coordinate inconsistencies, and spatially aggregated to a uniform 2 km × 2 km grid to ensure consistency across variables. This resolution minimizes empty-grid bias while preserving sufficient local detail, thus achieving a balance between meso-scale precision and computational efficiency, consistent with comparable urban retail studies listed in

Table 2. Data sources and variable descriptions.

Data Type	Variable Name	Description	Source
Retail POI Data	Retail location	Geographic coordinates and classification of retail stores	Amap (Gaode) Open Platform (Version 2023)
	Retail category	Reclassified into: food & beverage, apparel, electronics, supermarkets	National Economy Industry Classification Standard (GB/T 4754-2017)
Socioeconomic	Population density (POP)	Population per km2 in each 2 km × 2 km grid	Taiyuan Statistical Yearbook 2024; Census data
	School density (Sch)	Number of schools per km2	Taiyuan Municipal Education Bureau
	Hospital density (Hos)	Number of hospitals/clinics per km2	Taiyuan Health Commission; OpenMap
	Nighttime Light Intensity (NTL)	Average visible light radiance derived from VIIRS 2023 composite imagery, used as a proxy for economic activity	National Oceanic and Atmospheric Administration (NOAA)
	Housing price (Hou)	Average real estate price (CNY/m2) by community	Taiyuan Housing Market Report; Anjuke.com
	Company and enterprise density (Com)	Number of registered companies per km2	Fifth National Economic Census
Transport	Road density (Roa)	Total road length per km2	OpenStreetMap; Taiyuan Transport Bureau
	Transit station density (Sta)	Number of bus/subway stations per km2	OpenStreetMap; Amap (Gaode) Open Platform
	Parking lot density (Par)	Number of parking facilities per km2	OpenStreetMap; Amap (Gaode) Open Platform
Natural	Elevation (DEM)	Average elevation in meters per grid cell	ASTER Global Digital Elevation Model (GDEM)
	Slope (Slp)	Average slope in degrees per grid cell	Derived from DEM using ArcGIS Pro (Version 10.6) Spatial Analyst
	Terrain ruggedness (Flu)	Relative elevation difference within grid (ruggedness index)	Computed using ArcGIS Pro (Version 10.6) neighborhood analysis

[8,45]. In Table 2, the ‘Variable Name’ column lists each indicator together with its short name (in parentheses), which is used consistently in subsequent analyses.

Station density. Station density was calculated by aggregating all public transport nodes—including bus stops, subway stations, and high-speed rail stations—without weighting, consistent with the assumption that each node contributes equally to local accessibility. Point features were spatially joined to the 2 km × 2 km grid and converted to density values (count per km²).

Road density. Road density was derived from OpenStreetMap (OSM) polyline data. To avoid overrepresentation of minor footpaths and residential lanes (which account for over 80% of OSM road records in Taiyuan), only primary and secondary road classes were included. Total road length within each grid cell was summed and normalized by grid area (km/km²).

Company and enterprise density. Company POI data were obtained from Tianyancha. Duplicate records were removed based on name, address, and registration code. All enterprises—regardless of firm size—were treated equally because the GeoDetector method is designed to evaluate spatial heterogeneity rather than firm-specific economic capacity. Density was calculated as the number of enterprises per km².

Collinearity assessment. To contextualize GeoDetector results, pairwise Pearson correlations among the 12 independent variables were computed. Moderate correlations were observed for a few socioeconomic variables (e.g., population density vs. NTL), but no coefficients exceeded 0.80, indicating acceptable collinearity [48]. Moreover, GeoDetector’s q-statistic is theoretically robust to multicollinearity because the method does not require variable independence. The consistent q-ranking across stratification schemes further confirms that interaction detection and factor detection are not unduly influenced by variable correlations.

3.3. Spatial Analysis Methods

A multi-tier GIS-statistical framework was applied to explore spatial patterns and explanatory mechanisms, using methods validated in urban retail studies [10,21,22].

Average Nearest Neighbor (ANN) Analysis: This evaluates whether the spatial arrangement of retail outlets is clustered, random, or dispersed. The ANN ratio (R) compares the observed mean nearest neighbor distance with that expected under randomness (R = 1). Values of R < 1 indicate clustering, R ≈ 1 a random pattern, and R > 1 dispersion. This provides a preliminary test of overall spatial concentration before further local analyses.

Kernel Density Analysis: Reveals spatial continuity and commercial hotspots. The analysis utilized a fixed bandwidth (search radius) of 1.32 km. The selection of the 1.32 km bandwidth was based on a systematic sensitivity analysis to ensure appropriate smoothing for intra-urban patterning. We tested a range of alternative bandwidths (0.66 km, 1.32 km, and 2.64 km) to evaluate the stability and local accuracy of the resulting density surface. As documented in Appendix A (

Table A2. KDE Bandwidth Sensitivity Analysis (in kilometer).

Bandwidth (Kilometer)	Hotspot Pattern	Interpretation
0.66	Hotspots fragmented	Under-smoothed (unstable)
1.32	Clear, well-defined clusters	Optimal bandwidth used in main analysis
2.64	Over-smoothed; hotspots merge	Not suitable

and Figure A1), smaller bandwidths resulted in highly fragmented patterns, while larger bandwidths led to severe over-smoothing. The 1.32 km bandwidth provided the optimal balance between spatial smoothness and the capture of distinct local commercial centers. The 1.32 km search radius was applied to encompass the entire built-up area of Taiyuan to provide detailed spatial gradients within the 2 km × 2 km grid [24,42]. This parameter combination ensures a balance between spatial smoothness and local accuracy, consistent with previous KDE applications in medium-sized Chinese cities.

Edge Correction Acknowledgment: We note that a formal edge correction procedure was not applied during the KDE calculation. We acknowledge this as a limitation that may contribute to an underestimation of density near the study area boundaries. However, since the primary focus of our analysis is on the clustering patterns and explanatory mechanisms within the city’s established core built-up areas, the lack of correction is considered to have a minimal impact on the main conclusions regarding hotspot and cold spot identification. Future research should consider employing advanced boundary-weighted correction techniques.

Spatial Autocorrelation Analysis: Global Moran’s I and LISA indices identify overall clustering and local high/low-value clusters.

Location Quotient (LQ) Analysis: Assesses functional specialization of retail categories relative to city-wide averages. The index is defined as LQ = (R_i/R)/(T_i/T), where R_i and T_i represent the number of a given retail category and total outlets in a district, and R and T are their city-wide totals. LQ > 1 indicate local specialization, LQ = 1 parity, and LQ < 1 under-representation [29].

GeoDetector Analysis: Quantifies spatial explanatory power of natural, social, economic, and transportation factors. Interaction detection identifies bivariate enhancement or nonlinear amplification in spatial dynamics [8,18,49]. Its core statistical principle is the Q-statistic, which measures the degree to which factor X explains the spatial distribution of the dependent variable Y (retail density). The Q-statistic is calculated as:

$$Q=1-{\displaystyle \frac{{\sum }_{h=1}^L{N}_h{\mathsf{\sigma }}_h^2}{N{\mathsf{\sigma }}^2}}$$

(1)

where h = 1, …, L is the stratification of X, and ${\mathsf{\sigma }}^2$ and $\mathsf{\sigma }$ are the variances of Y in stratum h and the entire area, respectively. A Q-value close to 1 indicates that the factor X completely controls the spatial pattern of Y. This method is used to correlate with the individual and interactive explanatory power of the driving factors on retail agglomeration.

Variable stratification for GeoDetector: Because GeoDetector requires categorical strata for each explanatory variable, all continuous variables were discretized prior to q-statistic computation. Terrain-related variables (elevation, slope, ruggedness) were divided into five strata using the Jenks natural breaks method, which minimizes within-group variance. Socioeconomic and transportation variables (population density, school density, medical density, housing prices, NTL intensity, company and enterprise density, station density, road density, parking density) were partitioned into five equal-frequency (quantile) bins, following standard practice for urban intra-city heterogeneity analysis.

Significance testing: GeoDetector significance was assessed using 999-time permutations, and p-values < 0.001 were obtained for all factors. For robustness, 95% confidence intervals (CIs) were computed via bootstrapping.

Sensitivity to discretization: To evaluate binning sensitivity, we additionally tested two alternative schemes—(i) equal-interval binning and (ii) k-means clustering. The resulting q-rankings showed consistent top three factors (population density, NTL intensity, school density), confirming that the reported q-values are robust to classification method (Appendix A 

Table A3. Discretization Methods Used for GeoDetector.

Variable	Type	Discretization Method	Number of Bins
Elevation (DEM)	Natural	Jenks natural breaks	5
Slope	Natural	Jenks natural breaks	5
Terrain ruggedness	Natural	Jenks natural breaks	5
Population density	Social	Quantiles (equal-frequency)	5
School density	Social	Quantiles	5
Medical density	Social	Quantiles	5
Housing prices	Economic	Quantiles	5
NTL intensity	Economic	Quantiles	5
Company density	Economic	Quantiles	5
Station density	Transport	Quantiles	5
Road density	Transport	Quantiles	5
Parking density	Transport	Quantiles	5

).

All variables were normalized and rasterized to 2 km × 2 km units for consistency.

3.4. Research Process

The study follows a six-step workflow: data collection, preprocessing, grid construction, spatial pattern analysis, factor detection, and policy implication derivation. Figure 2 integrated approach ensures methodological rigor, spatial consistency, and theoretical-policy alignment [14,50,51].

4. Characterization of Spatial Distribution

4.1. Overall Spatial Patterns

Kernel density estimation (KDE) identifies Jinci Road (located south of Yingze Street and extending toward the Jinci scenic area in Jinyuan District, now functions as an emerging commercial axis) in Wanbailin District as a core high-density area, exhibiting a complex and mature structure best described as a double-axis plus multiple nodes pattern. This pattern along the Fen River, with additional hotspots in eastern Wanbailin, western Yingze, and northwestern Xiaodian. Peripheral districts, such as eastern Jinyuan and central Jiancaoping, show lower-density sub-centers due to terrain constraints (Figure 3). Average Nearest Neighbor (ANN) Analysis confirms strong clustering (R = 0.215, p-value < 0.01), consistent with synergistic urban commercial development (

Table 3. Average Nearest Neighbor analysis results for all retail outlets in Taiyuan.

Retail Formats	Average Nearest Neighbor Distance (d1, m)	Expected Average Nearest Neighbor Distance (d2, m)	Nearest Neighbors Index (R)	Z-Score	p-Value
All retail outlets	26.442	122.716	0.215	−219.223	0.000

). According to the Master Plan of Taiyuan City (1998–2010), the city’s traditional commercial system was dominated by a single city-level core centered on the Liuxiang–Zhonglou Street area, with Longtan, Chaoyang Street, Jiancaoping, and Xiayuan serving as supplementary sub-centers. The kernel-density and LISA analyses in this study reveal additional high-density retail clusters such as the Changfeng Business District, the South Railway Station area, and the Jinci Road corridor in Jinyuan District, indicating a transition from a historically monocentric structure to this established double-axis and multi-nodal spatial pattern. The core of this structure is a dual north–south axis framework formed along the Fen River valley, with functional activities primarily concentrated at multiple primary and secondary nodes along this framework (Figure 3).

Overall, Taiyuan’s retail network demonstrates a transition from monocentric to polycentric patterns, with a ‘double-axis plus multiple nodes’ structure centered on Jinci Road [30,51].

Z-score = standardized deviation from random expectation; p-value = significance level (<0.001 displayed as 0.000). All clusters are significant at p-value < 0.001.

It is important to note that the global ANN statistic assumes a homogeneous Poisson process (CSR) within the study window. However, retail outlets in Taiyuan exhibit strong spatial intensity variation, meaning that ANN captures only the overall deviation from spatial randomness and does not fully account for underlying inhomogeneity.

The extremely large Z-score (−219) obtained in this study is mathematically reasonable given (i) the very high number of POI points (over 21,000), (ii) the large study area (≈1400 km²), and (iii) the strong clustering of the retail system. Under such conditions, even small deviations in the nearest-neighbor distance from the CSR expectation relate to very large standardized values, which have been documented in other high-density urban POI datasets.

Although inhomogeneous K-function or network-constrained distance approaches may more accurately capture the clustering behavior along road networks, they require different methodological assumptions and were beyond the scope of this study. Therefore, ANN results here are interpreted only as an initial indication of global clustering, and finer-scale spatial patterns are further analyzed using KDE and LISA.

4.2. Analysis of Spatial Agglomeration and Location Advantages of Retail Sub-Industries

Retail outlets were classified into eight categories following GB/T 4754-2017: general retail, agriculture and food, daily apparel, graphics and textiles, medicine and healthcare, automotive power, digital appliances, and hardware/home furnishings (Figure 4).

General retail: This is distributed across Yingze, Xinghualing, Xiaodian, and Wanbailin, showing significant clustering (Moran’s I = 0.36, p < 0.01) and specialization in Xinghualing (LQ = 1.33) [29].

Agriculture and food: This is clustered in residential areas such as Yingze and Xinghualing (Moran’s I = 0.43; LQ = 1.49), reflecting proximity-driven consumption [31].

Daily apparel: This is concentrated in Liuxiang Business District (Yingze) with high specialization (LQ = 1.96).

Graphics and textiles: This is centered near educational hubs in Yingze and Xiaodian (Moran’s I = 0.31; LQ > 1.4), showing education–commerce integration [52].

Medicine and healthcare: There is high imbalance with Wanbailin as a hub (Moran’s I = 0.49; LQ = 2.06), while other districts lack specialization.

Digital appliances: This is concentrated along major traffic corridors in Yingze and Xiaodian (Moran’s I = 0.26; LQ > 1.6).

Automotive power: This is sparse and edge-oriented, favoring transport nodes; location advantages are in Xinghualing and Xiaodian (LQ > 1.1) [33].

Hardware and home furnishings: There is multipolar diffusion in Wanbailin, Jiancaoping, and Jinyuan (Moran’s I = 0.30; LQ > 1.6) [48].

Overall, retail sectors display spatial differentiation and specialization, with Yingze and Xiaodian as dominant cores, and Jiancaoping and Jinyuan lagging. To enhance clarity and conciseness in the presentation of LQ results,

Table 4. Dominant Retail Specialization by District (Top LQ Categories).

District	Dominant Retail Category	LQ	Interpretation
Yingze	Daily-use apparel	1.96	Strong specialization in fashion retail
Xinghualing	Agriculture and food	1.49	Food retail concentration in residential zones
Wanbailin	Medicine and healthcare	2.06	Healthcare facilities cluster
Xiaodian	Digital appliances	1.76	High-tech and electronics specialization
Jiancaoping	Hardware and home furnishings	1.64	Strong home-improvement orientation
Jinyuan	Hardware and home furnishings	1.61	Suburban hardware and construction retail focus

summarizes the dominant retail specializations for each district, highlighting the highest Location Quotient (LQ) categories. This concise presentation enables a quick overview of inter-district differences. The full results, including all retail categories and statistical indicators, are provided in Appendix A 

Table A4. Full statistical results of retail outlet count, proportions, and Location Quotient (LQ) by district and category in Taiyuan.

District	Retail Categories	Outlet Count	Proportion (%)	Location Quotient (LQ)	Specialization Interpretation
Xiaodian District	General retail	814	2.76	0.97	Balanced distribution
	Agriculture and food	570	1.93	1.04	Mild specialization in daily food services
	Daily-use apparel	751	2.55	1.32	Moderate specialization in apparel retail
	Graphics and textiles	201	0.68	1.48	Strong specialization near educational areas
	Medicine and healthcare	267	0.91	0.73	Below-average medical resource presence
	Automotive power	55	0.19	1.18	Slight concentration near peripheral roadways
	Digital appliances	236	0.80	1.76	High specialization in electronics retail
	Hardware and home furnishings	276	0.94	0.52	Underrepresented in home-related sectors
Wanbailin District	General retail	1312	4.30	0.85	General retail moderately dispersed
	Agriculture and food	914	3.00	0.91	Slightly underrepresented in food services
	Daily-use apparel	673	2.21	0.64	Weak apparel presence
	Graphics and textiles	164	0.54	0.66	Limited educational retail linkage
	Medicine and healthcare	1394	4.57	2.06	Strong specialization in healthcare facilities
	Automotive power	77	0.25	0.90	Moderate presence near transport corridors
	Digital appliances	169	0.55	0.69	Low concentration of electronics outlets
	Hardware and home furnishings	1123	3.68	1.15	Mild specialization in home improvement retail
Yingze District	General retail	827	7.07	0.98	Balanced and central general retail distribution
	Agriculture and food	549	4.69	1.00	Standard presence in food-related retail
	Daily-use apparel	1123	9.60	1.96	Strong specialization in apparel and fashion retail
	Graphics and textiles	241	2.06	1.76	Significant cluster of education-linked retail
	Medicine and healthcare	11	0.09	0.03	Severely underrepresented in medical retail
	Automotive power	31	0.26	0.66	Weak automotive service presence
	Digital appliances	215	1.84	1.60	Strong electronics retail specialization
	Hardware and home furnishings	192	1.64	0.36	Low specialization in home goods
Xinghualing District	General retail	1081	6.35	1.33	Clear specialization in general retail
	Agriculture and food	786	4.62	1.49	Strong specialization in residential food services
	Daily-use apparel	540	3.17	0.98	Balanced apparel retail presence
	Graphics and textiles	152	0.89	1.15	Slightly elevated educational retail linkage
	Medicine and healthcare	16	0.09	0.04	Very limited healthcare retail
	Automotive power	62	0.36	1.37	Specialization in edge-located vehicle services
	Digital appliances	101	0.59	0.78	Below-average electronics retail presence
	Hardware and home furnishings	333	1.96	0.65	Underdeveloped in home-related retail
Jiancaoping District	General retail	838	2.93	1.00	Typical general retail distribution
	Agriculture and food	454	1.59	0.83	Slight underrepresentation in food retail
	Daily-use apparel	456	1.60	0.80	Weak apparel market presence
	Graphics and textiles	87	0.30	0.64	Limited educational retail
	Medicine and healthcare	293	1.03	0.80	Below-average healthcare outlet density
	Automotive power	54	0.19	1.16	Peripheral vehicle services specialization
	Digital appliances	115	0.40	0.86	Modest electronics retail presence
	Hardware and home furnishings	869	3.04	1.64	Strong specialization in home improvement retail
Jinyuan District	General retail	773	2.67	1.00	Standard general retail distribution
	Agriculture and food	402	1.39	0.80	Underdeveloped food service sector
	Daily-use apparel	288	1.00	0.55	Weak specialization in apparel
	Graphics and textiles	70	0.24	0.56	Very low educational retail presence
	Medicine and healthcare	495	1.71	1.46	Moderate specialization in medical retail
	Automotive power	36	0.12	0.84	Modest automotive retail share
	Digital appliances	65	0.22	0.53	Sparse electronics distribution
	Hardware and home furnishings	784	2.71	1.61	High specialization in home & construction retail

for reference and reproducibility.

See Appendix A Table A4 for the complete dataset. Proportion refers to the percentage of each retail category within district totals; values do not sum to 100% because they represent within-district, not citywide, shares.

The Location Quotient (LQ) analysis is performed at the district level, which reflects administrative specialization rather than fine-scale spatial clustering. In contrast, the KDE and LISA analyses are conducted on the 2 km × 2 km grid, allowing intra-urban variations to be captured at a higher spatial resolution. The use of two complementary scales follows common practice in urban spatial-structure studies, where district-level indicators reflect macro-level functional differences while grid-level indicators describe local agglomeration patterns.

To ensure that the scale shift does not bias interpretation, we calculated a simple grid-level specialization index (the ratio of category-specific grid density to the citywide average). The spatial distribution of top-specialization grids closely matches the high-LQ districts, indicating that district-level specialization patterns are consistent with finer-scale variations.

In discussing the results, district-level LQ findings are therefore interpreted only at the administrative scale, and are not directly merged with grid-based LISA or KDE patterns in the same inference statement.

4.3. Spatial Autocorrelation Feature Recognition

Global Moran’s I indicates significant clustering across all retail categories (

Table 5. Global Moran Index and statistical test results.

Retail Formats	Global Moran’s I	Z-Score	p-Value	Description of the Space Model
All retail outlets	0.578	27.849	0.000	Overall significant spatial clustering distribution
General retail	0.593	28.625	0.000	Highly concentrated, showing the characteristics of multi-center layout
Agriculture and food	0.567	27.398	0.000	Significant clustering, tending to be distributed around residential areas
Daily-use apparel	0.395	20.156	0.000	Medium agglomeration, mainly in street-level commercial nodes
Graphics and textiles	0.472	23.422	0.000	Medium intensity agglomeration, favoring cultural or school neighborhoods
Medicine and healthcare	0.522	25.509	0.000	Stronger agglomeration, clearly dependent on large neighborhoods and hospital distribution
Automotive power	0.312	15.449	0.000	Weakly clustered, with distribution more dispersed along major transportation arteries
Digital appliances	0.390	21.491	0.000	Medium agglomeration, tends to be concentrated in large shopping areas or central districts
Hardware and home furnishings	0.190	9.711	0.000	Weak spatial agglomeration and loose layout

), while LISA identifies ‘high-high’ clusters in Yingze Business Circle, Changfeng Business District, and Proximity Street–South Central Ring Road, forming the main retail belt (Figure 5). ‘Low-low’ clusters appear in northern Jiancaoping and central-southern Jinyuan, confirming a pronounced north–south gradient. Sub-sector analyses reveal overlapping clusters for consumer-oriented industries (general retail, daily apparel, medicine), while ancillary and specialized sectors (agriculture–food, hardware, automotive, digital appliances) extend to suburban or transport-linked nodes.

The spatial autocorrelation analysis was conducted at the 2 km × 2 km grid level, with retail outlet density serving as the input variable for both Global Moran’s I and LISA, ensuring consistency with the scale of other spatial analyses in this study. Global Moran’s I values were computed using a row-standardized queen-contiguity spatial-weights matrix, which defines neighborhood relations by shared edges or vertices. Significance was assessed using 999 random permutations, and the associated Z-scores represent standardized deviations from random expectation.

For the LISA local indicators, statistical significance was first evaluated at p < 0.05, and then adjusted using the Benjamini–Hochberg False Discovery Rate (BH–FDR) procedure to control for multiple comparisons across the 407 grid cells. After BH–FDR adjustment (FDR ≤ 0.05), 191 significant local clusters remained (out of 194 pre-adjustment), indicating that the identified hotspot and cold spot patterns are robust to multiple-testing correction.

Grids with zero retail count were retained in the analysis rather than removed, as they contribute meaningful spatial contrast and are necessary for detecting low–low clusters and boundary transitions in local spatial autocorrelation. Values shown as 0.000 in Table 5 indicate p-values < 0.001, reflecting strong and statistically significant spatial clustering.

4.4. Preliminary Study on Factors Influencing Spatial Patterns

A Pearson’s correlation analysis was conducted between 12 natural, social, economic, and transportation factors and total retail distribution to identify linear relationships and key drivers of spatial clustering (

Table 6. Statistical distribution of retail outlets by classification intervals of impact factors.

Factor Type	Classification Interval/Range Interval	Class 1	Class 2	Class 3	Class 4	Class 5
DEM/m	Elevation Interval (m)	683–855	855–1011	1011–1182	1182–1359	1359–1816
	Area Proportion (%)	49.57	16.12	12.83	14.81	6.68
	Retail Outlets Proportion (%)	95.04	4.91	0.04	0.01	0
Slope/°	Slope Interval (°)	0–4.89	4.89–10.51	10.51–17.35	17.35–26.88	26.88–62.31
	Area Proportion (%)	46.9	23.38	17.09	10.21	2.42
	Retail Outlets Proportion (%)	74.7	22.17	2.81	0.28	0.05
Fluctuation/m	Terrain Ruggedness Interval (m)	0–8	8–17	17–29	29–48	48–154
	Area Proportion (%)	53.73	23.29	15.97	6.03	0.97
	Retail Outlets Proportion (%)	84.76	13.89	1.2	0.14	0.01
POP/(103 person/km2)	Population Density Interval	0.21–22.31	22.31–65.51	65.51–114.74	114.74–179.04	179.04–256.40
	Number of Outlets	4382	4501	3003	4771	4678
	Outlet Density (unit/km2)	3.91	28.64	51.57	105.81	162.24
Sch/(unit/km2)	School Density Interval (unit/km2)	0–1	1–2	2–3	3–7	7–13
	Number of Outlets	2488	2960	599	11,216	4072
	Outlet Density (unit/km2)	0.52	1.38	2.4	4.98	9.24
Hos/(unit/km2)	Hospital Density Interval (unit/km2)	0–1	1–5	5–12	12–25	25–47
	Number of Outlets	8667	3098	2533	3772	3265
	Outlet Density (unit/km2)	0.09	2.31	7.39	18.76	35.84
NTL/(nW/cm−2/sr−1/km2)	NTL Density Interval (nW/cm−2/sr−1/km2)	0–7.77	7.77–20.20	20.20–36.85	36.85–59.86	59.86–118.69
	Number of Outlets	335	1149	1695	4651	13,503
	Outlet Density (unit/km2)	0.47	4.71	10.95	27.75	93.78
Hou/(103 CNY/km2)	Housing Price Interval (103 CNY/km2)	3.25–5.23	5.23–6.97	6.97–8.94	8.94–14.16	14.16–27.18
	Number of Outlets	4572	2921	7749	5198	895
	Outlet Density (unit/km2)	4.43	47.85	52.51	36.42	27.97
Com/(unit/km2)	company and enterprise Density Interval (unit/km2)	0–12	12–40	40–94	94–188	188–387
	Number of Outlets	903	3928	3620	9027	3857
	Outlet Density (unit/km2)	1.1	10.9	45.59	83.59	80.35
Distance to Roads	Distance Interval (m)	≤500	≤1000	≤1500	≤2000	>2000
	Number of Outlets	13,557	15,064	15,466	20,669	21,335
	Outlet Proportion (%)	63.54	70.61	72.49	96.88	100
Par/(unit/km2)	Parking Density Interval (unit/km2)	0–1.12	1.12–2.75	2.75–7.50	7.50–15.00	15.00–42.50
	Number of Outlets	15,361	1015	1629	1601	1729
	Outlet Density (unit/km2)	11.66	25.38	104.22	50.05	144.08
Sta/(unit/km2)	Station Density Interval (unit/km2)	0–0.36	0.36–1.11	1.11–2.25	2.25–4.50	4.50–8.26
	Number of Outlets	836	663	1703	7082	11,051
	Outlet Density (unit/km2)	1.29	2.49	8.4	40.8	89.13

). Before conducting the analysis, Pearson correlation was framed in terms of identifying linear associations under approximate homoscedasticity, based on visual inspection of scatterplots and residual–fitted patterns, rather than relying on normality assumptions. To ensure robustness with respect to monotonic but potentially non-linear relationships, Spearman rank correlations were also computed. The complete Pearson and Spearman correlation matrices, together with variable-wise sample sizes based on pairwise non-missing observations, are provided in Appendix A 

Table A5. Pearson correlation matrix for natural, socioeconomic, and transportation variables.

	DEM	Slo	Flu	Pop	Sch	Hos	NTL	Pri	Com	Roa	Par	Sta
DEM	1	0.76	0.53	−0.32	−0.36	−0.22	−0.56	−0.07	−0.35	−0.39	−0.12	−0.44
Slo	0.76	1	0.79	−0.33	−0.37	−0.2	−0.58	−0.08	−0.37	−0.38	−0.14	−0.47
Flu	0.53	0.79	1	−0.23	−0.27	−0.06	−0.43	−0.1	−0.29	−0.24	−0.15	−0.35
Pop	−0.32	−0.33	−0.23	1	0.81	0.3	0.78	0.1	0.67	0.48	0.39	0.78
Sch	−0.36	−0.37	−0.27	1	1	0.44	0.76	0.28	0.71	0.49	0.31	0.8
Hos	−0.22	−0.2	−0.06	0.3	0.44	1	0.51	−0.03	0.42	0.45	0.03	0.49
NTL	−0.56	−0.58	−0.43	0.78	0.76	0.51	1	0.09	0.7	0.66	0.29	0.81
Pri	−0.07	−0.08	−0.1	0.1	0.28	−0.03	0.09	1	0.16	0.27	0.1	0.07
Com	−0.35	−0.37	−0.29	0.67	0.71	0.42	0.7	0.16	1	0.43	0.3	0.72
Roa	−0.39	−0.38	−0.24	0.48	0.49	0.45	0.66	0.27	0.43	1	0.22	0.58
Par	−0.12	−0.14	−0.15	0.39	0.31	0.03	0.29	0.1	0.3	0.22	1	0.46
Sta	−0.44	−0.47	−0.35	0.78	0.8	0.49	0.81	0.07	0.72	0.58	0.46	1

and

Table A6. Spearman rank correlation matrix for natural, socioeconomic, and transportation variables.

	DEM	Slo	Flu	Pop	Sch	Hos	NTL	Pri	Com	Roa	Par	Sta
DEM	1	0.86	0.62	−0.5	−0.55	−0.44	−0.74	−0.22	−0.7	−0.42	−0.13	−0.52
Slo	0.86	1	0.81	−0.55	−0.6	−0.4	−0.76	−0.2	−0.75	−0.42	−0.21	−0.59
Flu	0.62	0.81	1	−0.41	−0.44	−0.18	−0.61	−0.1	−0.56	−0.27	−0.23	−0.47
Pop	−0.5	−0.55	−0.41	1	0.7	0.34	0.86	0.03	0.76	0.55	0.36	0.72
Sch	−0.55	−0.6	−0.44	0.7	1	0.51	0.81	0.11	0.83	0.53	0.36	0.82
Hos	−0.44	−0.4	−0.18	0.34	0.51	1	0.48	0.02	0.46	0.4	0.02	0.42
NTL	−0.74	−0.76	−0.61	0.86	0.81	0.48	1	0.21	0.88	0.61	0.33	0.8
Pri	−0.22	−0.2	−0.1	0.03	0.11	0.02	0.21	1	0.3	0.27	−0.03	0.17
Com	−0.7	−0.75	−0.56	0.76	0.83	0.46	0.88	0.3	1	0.56	0.35	0.82
Roa	−0.42	−0.42	−0.27	0.55	0.53	0.4	0.61	0.27	0.56	1	0.24	0.53
Par	−0.13	−0.21	−0.23	0.36	0.36	0.02	0.33	−0.03	0.35	0.24	1	0.37
Sta	−0.52	−0.59	−0.47	0.72	0.82	0.42	0.8	0.17	0.82	0.53	0.37	1

.

Prior to testing, all variables were normalized using the min–max method, and normality was confirmed using the Kolmogorov–Smirnov test (Equation (2)):

$$X_{\mathrm{norm}}={\displaystyle \frac{X-X_{\min }}{X_{\max }-X_{\min }}}$$

(2)

where X_norm is the normalized value, X the original variable, and X_min, X_max the minimum and maximum values in the dataset for input into the GeoDetector framework.

Continuous variables such as elevation, slope, and terrain ruggedness were categorized using the natural-breaks (Jenks) classification method, which minimizes within-group variance and maximizes between-group variance. The thresholds (e.g., elevation < 855 m, slope < 10.51°, ruggedness < 17) therefore correspond to statistically derived class breakpoints rather than arbitrary cut-offs. These procedures ensured that the dataset met the assumptions required for correlation and GeoDetector analyses.

Natural Factors: Elevation (<855 m), slope (<10.51°), and terrain ruggedness (<17 m) constrain retail location, concentrating most outlets in accessible terrain (r values: +0.98, −0.98, −0.92, p < 0.01). The high inter-correlation among these natural factors (|r| ≈ 0.76–0.86; Appendix A Table A5) reflects their shared derivation from DEM-based surfaces, which naturally produces strong covariation. While such multicollinearity may affect regression-based models, it does not bias GeoDetector results, as the q-statistic does not require independent explanatory variables.

Social Factors: Population density (r = 0.77), school density (r = 0.67), and hospital density (r = 0.54) strongly influence clustering.

Economic Factors: Nighttime light (proxy for GDP, r = 0.67), housing prices (r = 0.39), and company and enterprise density (r = 0.52) indicate consumption capacity and industrial linkage effects [8,35,37,38,39].

Transportation Factors: Station density shows strong positive correlation (r = 0.71); road density (r = 0.36) and parking density (r = 0.31) moderately promote clustering.

Black points denote all retail outlets; their high density may partially obscure background layers at smaller scales. A composite multi-factor overlay index was then generated by normalizing and summing the spatial layers of population density, station density, nighttime light intensity, road buffer accessibility, and topographic elevation. The resulting values were rescaled to a 0–1 range, where 0 indicates minimal combined influence (low population, weak accessibility, low economic activity, or unfavorable terrain), and 1 indicates maximum combined influence reflecting strong synergies among key drivers of retail clustering. This integrated surface is visualized in Figure 6f Multi-factor overlay.

Spatial overlays of population, stations, NTL, roads, and elevation reinforce synergistic effects in core districts (Yingze, Changfeng) and illustrate constraints in ‘low-low’ cold spots (Jiancaoping, Jinyuan). These findings provide a foundation for quantitative GeoDetector analysis of factor interactions in Section 5 (Figure 6).

5. Drivers of Retail Spatial Distribution

This section quantitatively explores the determinants of retail outlet spatial distribution in Taiyuan, based on GeoDetector analysis. Building on the pattern characterization in Section 4, the model evaluates the explanatory power of 12 variables (covering socioeconomic, natural, and transportation dimensions) using outlet distribution density as the dependent variable. Additionally, the interaction detector identifies compound effects and nonlinear synergies among factors, revealing underlying mechanisms of spatial configuration.

5.1. Factor Construction and Data Processing

The analysis adopts 2 km × 2 km grids across six districts as spatial units for 2023. Guided by prior literature and Section 4 findings, twelve variables were selected across four dimensions:

5.1.1. Natural Factors

Elevation (DEM): extracted from ASTER GDEM data;

Slope: derived from DEM data;

Terrain ruggedness: represented by the standard deviation of the mean elevation values within each 2 km × 2 km grid cell, indicating local topographic variation.

5.1.2. Social Factors

Population density: from Taiyuan’s Seventh Census;

School density: based on POI data for all educational levels;

Medical facility density: derived from POI data of hospitals and clinics.

5.1.3. Economic Factors

NTL density (used as a proxy for overall urban vitality): sourced from the Visible Infrared Imaging Radiometer Suite (VIIRS) dataset. Compared with conventional GDP statistics, which in the Taiyuan Statistical Yearbook are only available at the district level and cannot capture intra-urban heterogeneity, NTL data provide spatially continuous and fine-resolution information. Therefore, NTL serves as an effective substitute for grid-level economic activity, reflecting overall human activity intensity and development levels rather than solely nighttime economic behavior [45].

Housing prices: 2023 averages from real estate platforms.

Company and enterprise density: represents the total number of registered companies and enterprises per square kilometer, derived from the Tianyancha business registry database.

5.1.4. Transportation Factors

Road density: computed using OpenStreetMap vectors;

Parking lot density: from POI sources;

Transportation station density: including bus, subway, and high-speed rail.

All variables were normalized using the min–max method (ArcGIS Pro Version 10.6).

5.2. Single-Factor Detection Results

The Factor Detector quantifies each variable’s explanatory strength via q-values (0–1), capturing spatial stratified heterogeneity beyond linear correlation [38].

Table 7. Ranking of Factor Detection q-Values.

Dimension	Factor Name	q-Value	Rank
Social	Population density	0.59	1
Economic	NTL density	0.51	2
Social	School density	0.50	3
Natural	Elevation (DEM)	0.44	4
Transportation	Transportation station density	0.42	5
Economic	Housing price	0.39	6
Social	Medical facility density	0.32	7
Transportation	Road density	0.29	8
Natural	Slope	0.27	9
Transportation	Parking lot density	0.26	10
Natural	Terrain ruggedness	0.24	11
Economic	Company and enterprise density	0.23	12

summarizes the ranked q-values.

Robustness checks based on alternative stratification schemes (quantile, equal-interval, and k-means) yielded stable q-rankings, indicating that the factor detection results are not sensitive to binning choices. All q-values were statistically significant at p < 0.001 with narrow 95% confidence intervals.

Population density (q = 0.59) emerges as the most significant predictor, emphasizing the role of demographic agglomeration. NTL and school density also rank highly, reflecting the combined effects of urban vitality, consumption potential, and commercial clustering. Natural factors and transportation variables contribute moderately, while terrain ruggedness and Company and enterprise density exert limited direct influence [36].

5.3. Interaction Detection Results

The Interaction Detector assesses the joint explanatory power of factor pairs, classifying interactions as bivariate enhancement, weak enhancement, or no interaction [8]. Key results are shown in

Table 8. Selected Interaction Detection Results.

Factor A	Factor B	Combined q-Value	Interaction Type
Population Density	NTL density	0.72	Bivariate enhancement
School Density	Transportation Station Density	0.68	Bivariate enhancement
Elevation (DEM)	Road Density	0.61	Bivariate enhancement
Housing Price	Company and Enterprise Density	0.54	Weak enhancement
Slope	Terrain Ruggedness	0.37	No interaction

.

Most factor pairs exhibit synergistic effects. For example, the combination of population and NTL density (q = 0.72) strongly enhances retail clustering, highlighting the dual influence of demographic concentration and nighttime economic activity. This utilization of NTL is well-supported, as NTL intensity serves as a reliable, high-resolution proxy for assessing urban economic vibrancy, especially the nocturnal consumption and service intensity [45,53]. Similarly, school and station density reinforce each other (q = 0.68), suggesting that educational hubs near transport nodes act as commercial magnets. Ruggedness-related variables show negligible interaction, indicating independent and limited effects. These results provide the basis for the multidimensional mechanisms outlined below.

5.4. Multi-Dimensional Mechanisms of Influence

Combining GeoDetector outcomes with spatial correlation patterns, the following five dimensions shape retail spatial organization:

5.4.1. Socio-Demographic Foundations

High-density population zones (e.g., Yingze, Xiaodian) serve as primary retail anchors, sustaining outlet concentrations exceeding 100 units/km². This reflects central place theory, where consumer thresholds define viable commercial centers [11,36].

5.4.2. Economic Vitality and Format Upgrading

Areas like Changfeng and Qinxian, with NTL above 80,000 nW/cm⁻²/sr⁻¹/km², exhibit premium retail agglomerations (up to 93.78 units/km²), in line with economic base theory [54].

5.4.3. Social Infrastructure as Catalysts

School-dense zones (7–13 schools/km²) host 68.6% of outlets; medical clusters (25–47 institutions/km²) support 35.84 units/km², illustrating a ‘service multiplier effect’ [34].

5.4.4. Natural Constraints

Over 50% of the study area lies above 855 m or on slopes > 10.5°, yet these zones accommodate only 4.96% of outlets—confirming topography as a limiting factor.

5.4.5. Transportation Accessibility

Road-proximate grids (within 500 m) hold 63.54% of outlets, while subway-rich zones (4.5–8.26 stations/km²) support up to 89.13 units/km². This confirms the gravity model of spatial interaction [36,51].

5.4.6. Institutional and Planning Factors

Beyond natural, socioeconomic, and transportation determinants, Taiyuan’s retail spatial configuration is also strongly influenced by institutional and planning frameworks. The Taiyuan announces 2021–2035 land and space plan and the Taiyuan’s 14th Five-Year Plan for Commercial Development promote balanced growth between the old city and emerging districts. Renewal programs in Yingze and Xinghualing focus on upgrading historical commercial cores, while industrial-to-service transformation in Wanbailin and Xiaodian fosters new sub-centers such as Changfeng Business District and South Railway Station. These policies have guided the redistribution of retail activities along renewal corridors and transit axes, highlighting the interaction between government planning and market dynamics in shaping Taiyuan’s polycentric retail landscape (Figure 7).

6. Conclusions and Research Limitations

6.1. Research Conclusions and Policy Implications

This study, based on 2023 POI data and GIS spatial analysis, reveals the spatial distribution patterns and driving mechanisms of retail outlets in Taiyuan’s main urban area. Results demonstrate that the retail structure exhibits a polycentric and road-dependent configuration, with significant clustering in Yingze and Xiaodian districts, while peripheral areas such as Jiancaoping and Jinyuan remain under-served. Retail formats show spatial differentiation, reflecting both functional specialization and uneven service provision. Among the explanatory variables, population density, nighttime light intensity, and school distribution exert the strongest influence, while topographic factors impose spatial constraints on expansion.

From a policy perspective, the findings suggest the need for (1) promoting a polycentric commercial system by cultivating secondary retail centers in densely populated, well-connected districts; (2) improving retail service equity in peripheral zones through targeted infrastructure and land-use planning; and (3) integrating transportation hubs with retail facility planning to enhance accessibility. These strategies contribute to the optimization of spatial equity and resource allocation in transition cities undergoing rapid restructuring.

Beyond general policy implications, the empirical findings of this study provide direct support for spatial decision-making in retail planning and urban governance. First, the grid-level identification of high- and low-supply retail zones offers a quantitative basis for prioritizing land-use allocation and commercial facility deployment, especially in peripheral districts such as Jiancaoping and Jinyuan. Second, the interaction effects revealed by the GeoDetector model, particularly the synergistic impacts of population density, nighttime light, and transport accessibility, enable planners to evaluate where commercial vitality can be most effectively strengthened under limited land resources. Third, the multi-factor overlay results can be incorporated into municipal spatial-planning platforms as a decision tool for simulating development scenarios, monitoring service inequality, and guiding differentiated planning at the core–subcenter–periphery scale. Finally, by linking retail distribution with indicators of community service equity, the study provides a practical analytical framework that supports evidence-based commercial restructuring and enhances the spatial efficiency of public-service provision in transition cities.

In practical terms, the optimization of Taiyuan’s retail spatial layout can proceed through three levels of implementation. First, establish differentiated strategies for core, sub-center, and peripheral zones to encourage functional diversification and public-space enhancement in central districts (e.g., Yingze), while developing neighborhood-scale retail in emerging sub-centers (e.g., Wanbailin and Xiaodian). Second, integrate retail distribution with ‘15-min living circle’ planning by identifying underserved areas through grid-based accessibility mapping to ensure service equity. Third, promote the adaptive reuse of underutilized industrial parcels for new-format retail clusters, such as cultural and creative spaces. These actions are expected to improve retail accessibility, stimulate urban vitality, and foster balanced commercial development across Taiyuan’s main districts.

6.2. Research Innovations and Theoretical Contributions

This study makes three main contributions. First, it demonstrates the value of high-resolution POI data integrated with a GIS-based framework for analyzing retail spatial structures at the intra-urban scale, thereby advancing methodological approaches in urban geography and spatial information science. Second, by applying the GeoDetector model to capture both individual and interactive effects of multidimensional factors, the study introduces a robust quantitative approach that extends beyond linear correlation methods. Third, the focus on Taiyuan, a resource-based transition city, expands the empirical scope of retail spatial studies, offering generalizable insights for cities facing industrial restructuring and spatial inequality challenges.

Unlike studies centered on coastal service-oriented metropolises such as Shanghai or Guangzhou, this research focuses on a resource-based transition city characterized by industrial restructuring and population redistribution. The spatial evolution observed in Taiyuan, where retail clusters emerge along renewal corridors and converted industrial land. This differs from service cities dominated by tertiary-sector concentration or tourism cities shaped by seasonal flows (e.g., Xi’an and Luoyang). This distinction underscores the role of industrial transformation and institutional renewal in defining commercial spatial patterns, thus extending the theoretical applicability of retail geography to resource-transition contexts.

Together, these contributions highlight the methodological and empirical significance of combining spatial data with GIS-based modeling, providing a transferable framework for retail spatial analysis in diverse urban contexts.

6.3. Research Limitations and Future Prospects

Despite efforts to improve data accuracy and analytical dimensions, this study has limitations. First, the use of 2023 POI data limits longitudinal analysis. Future work could incorporate multi-temporal data to reveal retail spatial distribution. Second, POI classification precision remains imperfect due to ambiguous categories. Enhancing this with tax or business registration data could improve clustering accuracy. Third, key non-spatial factors (e.g., land policies, consumer preferences) were omitted; incorporating survey or trajectory data could address this. Fourth, the fixed 2 km × 2 km grid effectively captures intra-urban heterogeneity at a meso-scale suitable for citywide analysis. However, it may still overlook finer neighborhood-level patterns, which future multi-scale studies could address. Finally, although this study discusses retail service equity, it does not incorporate a formal accessibility model such as the two-step floating catchment area (2SFCA) method or network-based coverage analysis. The absence of an explicit accessibility measure represents a methodological limitation and should be addressed in future research.

In summary, this research offers empirical and theoretical insights into urban retail spatial organization in central China. Future studies should expand variable dimensions, refine classification methods, and enhance theoretical integration to better support equitable and efficient retail planning and governance.

7. Discussion

This integration of POI high-resolution spatial data and GIS-based methods provides new opportunities to examine the spatial dynamics of urban retail systems with higher precision and broader applicability. This study illustrates how kernel density estimation, spatial autocorrelation, and Location Quotient (LQ) can systematically identify clustering tendencies and specialization, while GeoDetector enables robust evaluation of the multidimensional drivers of retail spatial organization. Compared with conventional regression-based approaches, the adopted framework better captures spatial heterogeneity and nonlinear interactions, which are increasingly critical in understanding urban commercial distribution.

The empirical findings from Taiyuan reaffirm classical urban theories while extending them through high-resolution spatial data. The dominance of population density and accessibility factors aligns with Christaller’s central place theory and Tobler’s first law of geography, whereas reflects the ongoing polycentric restructuring typical of transition cities. Importantly, the study demonstrates that GIS-based approaches can effectively link spatial inequality with service equity, providing evidence-based support for urban policy design.

Beyond Taiyuan, the methodological framework has broad transferability. The integration of POI data, spatial statistical models, and GeoDetector can be applied to other resource-based cities (e.g., Datong, Shenyang, or Lanzhou) and even to non-retail urban functions such as healthcare, education, or cultural services. This adaptability underscores the role of GIS-based approaches as a universal analytical toolkit for urban spatial research, bridging academic inquiry with policy and planning practices.

Nevertheless, certain limitations remain. The study focuses primarily on spatial determinants, while consumer behavior and institutional factors are not directly addressed. Future work should incorporate multi-temporal POI datasets, behavioral survey data, or mobile trajectory data to more accurately capture temporal dynamics and demand-side variations. Moreover, integrating policy instruments and governance variables into spatial models would enhance the explanatory depth of urban retail studies.

In sum, this study contributes to spatial information science by providing an integrated and scalable framework for analyzing retail distribution. It validates the effectiveness of GIS-based methods in uncovering complex spatial mechanisms, highlights the significance of socio-demographic and infrastructural drivers, and demonstrates the transferability of the approach across urban contexts.