Temporal Evolution of Multi-Dimensional Built Environment Perceptions and Street Vitality: A Longitudinal Analysis in Rapidly Urbanizing Cities

Abstract

Rapid urbanization fundamentally transforms how residents perceive and interact with built environments, yet the dynamic relationships between these evolving perceptions and street vitality remain inadequately understood. As cities undergo rapid transformation, traditional assumptions about fixed perception–vitality relationships may no longer hold, necessitating a deeper understanding of how these relationships evolve over time and space. This study aims to investigate how multiple dimensions of built environment perception influence street vitality and how these relationships evolve spatially and temporally in rapidly urbanizing contexts. We developed a multi-level interpretative framework combining Multi-scale Geographically Weighted Regression (MGWR) with machine-learning-based SHAP analysis to analyze multi-source data from Hohhot, China, spanning 2019–2023. Our approach examined four key perception dimensions—comfort, safety, convenience, and pleasure—and their impacts on street vitality patterns during a period of intensive urban development. The analysis reveals three major findings: first, perception–vitality relationships evolved from highly heterogeneous spatial patterns toward increasing homogenization over time, suggesting urban development standardization effects driven by rapid urbanization processes. Second, several perception dimensions underwent significant transformations, with safety perception shifting from negative to positive influence and convenience perception displaying complex nonlinear threshold effects as urban infrastructure matured. Third, the relative importance of perception dimensions changed over time, reflecting evolving urban priorities and resident expectations shaped by urbanization experiences. These findings demonstrate that perception–vitality relationships are dynamic rather than static, challenging assumptions about fixed environmental effects in urban planning. The study provides empirical evidence for implementing adaptive, context-sensitive urban interventions that acknowledge both spatial heterogeneity and temporal evolution, offering valuable insights for enhancing street vitality in rapidly urbanizing environments worldwide.

1. Introduction

Street vitality, as a fundamental indicator of urban quality, has become increasingly critical amid rapid global urbanization [1,2]. The unprecedented pace of urbanization in developing countries, particularly China, has fundamentally transformed traditional street life patterns, creating complex urban environments where identical physical features may generate vastly different levels of vitality across spatial and temporal dimensions [3,4]. This transformation challenges conventional approaches to understanding urban vitality and necessitates new frameworks for examining the dynamic relationships between built environments and human activity patterns.

The theoretical foundation for street vitality research has evolved significantly over recent decades, drawing from multiple disciplinary streams that can be integrated through a coherent “perception–behavior–vitality” framework. This framework provides a logical mechanistic pathway connecting environmental psychology, behavioral geography, and urban vitality theories. At the foundational level, environmental psychology establishes that built environment characteristics are filtered through human perceptual processes, creating subjective evaluations of comfort, safety, convenience, and pleasure [5,6]. These perceptions then influence behavioral responses—spatial choices, activity patterns, and social interactions—as demonstrated in behavioral geography and space syntax research [7,8]. Finally, the aggregation of individual behavioral responses manifests as observable urban vitality patterns, connecting to street life theories and contemporary urban analytics approaches [9,10,11].

While foundational urban theories have long recognized the importance of street life and human-scale environments, this integrated framework enables more systematic understanding of the mechanistic pathways through which built environments generate vitality. Recent studies utilizing mobile phone data [9,10] and social media check-ins [11] have provided empirical evidence of spatial heterogeneity in vitality patterns, while advances in computer vision have enabled large-scale assessment of environmental perceptions [7,8], creating opportunities for comprehensive analysis across the perception–behavior–vitality pathway. Methodologically, local regression techniques like Geographically Weighted Regression (GWR) have emerged as valuable tools for analyzing spatial relationships [12], while machine learning interpretation methods, particularly SHapley Additive exPlanations (SHAP) values, offer approaches for understanding complex nonlinear relationships within this framework [13]. However, the integration of these methodologies for longitudinal analysis of perception–behavior–vitality pathways remains underexplored, particularly in rapidly urbanizing contexts where physical environments, perceptual filters, and behavioral patterns undergo continuous evolution.

Despite significant progress in street vitality research, three critical challenges limit our current understanding of perception–vitality relationships. First, the widespread adoption of global statistical models assumes spatial stationarity [14], failing to capture how identical built environment features may generate varying levels of vitality across different urban contexts. While recent studies utilizing big data have revealed significant spatial variations in vitality patterns [15,16], most analyses cannot explain why these variations occur or how they evolve over time, particularly in Chinese cities where rapid expansion creates diverse urban fabrics within single urban areas [17,18]. Second, current research demonstrates a predominant focus on objective physical measurements while insufficiently addressing the crucial role of human perception [19,20]. Although environmental psychology research has established perception’s significant influence on behavior [21], and computer vision advances have enabled large-scale perceptual assessment of urban environments [22,23], studies examining how these perceptions vary spatially across urban contexts or how their influence changes over time remain relatively limited [24]. This represents a significant gap in understanding the dynamic nature of perception–vitality relationships as urban environments mature and resident expectations evolve. Third, the interpretability of spatial variations in perception–vitality relationships remains largely unexplored [25]. While local regression techniques have found applications in urban studies [12], their use in examining perception–vitality relationships across different spatial scales and temporal periods is limited. The potential of machine learning interpretation techniques, particularly SHAP values, for understanding complex spatio-temporal relationships in longitudinal urban vitality analysis remains largely untapped [13].

The case of Hohhot, the capital city of Inner Mongolia Autonomous Region, provides an ideal context for addressing these research challenges while expanding the geographical diversity of street vitality research. As a representative northern grassland city of China, Hohhot embodies distinctive characteristics that differentiate it from extensively studied southern Chinese cities (such as Shenzhen and Shanghai) and Western urban centers that dominate current vitality research [26,27,28,29,30]. The city’s unique integration of Mongolian cultural heritage with rapid modernization, its semi-arid continental climate, and its distinct socio-spatial structure creates a valuable laboratory for examining perception–vitality relationships in underrepresented urban contexts. Furthermore, Hohhot’s urban development trajectory exemplifies the challenges facing many rapidly urbanizing Chinese cities. Its expansion has prioritized vehicular efficiency over human-scale considerations, resulting in a persistent trend of weakened public space vitality despite similar urban forms across different districts [31,32]. This phenomenon underscores the need for a more nuanced understanding of how environmental perceptions influence street vitality across different urban contexts and temporal periods, making it an exemplary case for longitudinal perception–vitality analysis.

To address these identified challenges, this study aims to investigate how multiple dimensions of built environment perception influence street vitality and how these relationships evolve spatially and temporally in rapidly urbanizing contexts. We propose a novel multi-level interpretative framework combining Multi-scale Geographically Weighted Regression (MGWR) with SHAP values to: (1) quantify spatial heterogeneity in perception–vitality relationships across different scales, (2) reveal varying importance of perceptual dimensions across space and time, and (3) trace the temporal evolution of these relationships over a critical three-year period (2019–2023) that encompasses significant urban development in Hohhot [33,34]. This longitudinal approach enables investigation of how perception–vitality relationships mature and stabilize as cities develop, providing temporal insights that complement existing cross-sectional studies [35]. The combination of local regression methods with interpretable machine learning techniques provides opportunities for examining how and why built environment–vitality relationships vary across space and time [36,37], contributing to understanding of evolutionary processes in urban environments [38].

Through this integrated spatio-temporal approach, this research aims to advance both theoretical understanding and practical applications in urban design. The study provides empirical evidence for developing context-sensitive and temporally adaptive interventions that can effectively enhance street vitality while acknowledging both spatial heterogeneity and temporal evolution of urban environments. The findings contribute to creating more vibrant, livable urban spaces that respond to local contexts while supporting sustainable urban development, with particular relevance for northern Chinese cities and similar urban environments that have been underrepresented in the vitality literature. The methodological contribution lies in demonstrating how MGWR and SHAP approaches can be integrated to capture both spatial heterogeneity and nonlinear effects while maintaining interpretability for practical urban planning applications. This framework enables both detailed local analysis and broader pattern recognition, supporting the development of nuanced, context-sensitive urban design strategies that acknowledge the dynamic nature of perception–vitality relationships in rapidly changing urban environments.

2. Data and Methods

This study employs an integrated methodological framework combining multi-source data analysis and advanced spatial–statistical techniques to investigate the spatial heterogeneity and temporal evolution of perception–vitality relationships (see Figure 1). Our longitudinal approach utilizes urban data collected across three time points (2019, 2021, and 2023) from Hohhot, China, including road network data, mobile phone signaling data, Point of Interest (POI) data, and street view imagery. These data sources underwent rigorous processing protocols to ensure quality, completeness, and cross-temporal consistency. Methodologically, we implemented a three-stage analytical process: (1) constructing comprehensive indices for street vitality and built environment perceptions through multi-dimensional measurements, (2) employing Multi-scale Geographically Weighted Regression (MGWR) to quantify spatial heterogeneity in perception–vitality relationships at varying spatial scales, and (3) utilizing eXtreme Gradient Boosting (XGBoost) machine learning with SHAP values to reveal complex nonlinear relationships and threshold effects. This integrated approach enables us to capture both spatial variations and nonlinear interactions while maintaining interpretability for practical urban planning applications, offering deeper insights into how perception–vitality relationships evolve spatially and temporally across the urban landscape.

2.1. Study Area

Hohhot, the capital city of Inner Mongolia Autonomous Region, serves as an ideal case study for examining spatial heterogeneity in street vitality for several reasons. First, as a typical rapidly developing Chinese city, Hohhot has experienced significant urban transformation over the past decades, with its urban built-up area expanding from 138 km² in 2000 to 243 km² in 2020, and its urban population growing from 1.12 million to 2.35 million. This rapid development provides a unique context for studying how built environment perceptions influence street vitality during urban transformation.

The study area focuses on the central urban area of Hohhot (111°35′0″ E–111°45′0″ E, 40°45′0″ N-40°50′0″ N), encompassing approximately 218 km² of urban core area. As shown in Figure 2, this area is characterized by diverse urban fabrics, including the historical old town centered around Dazhao Temple (dating back to the Qing Dynasty), the modern commercial districts along Zhongshan Road and Xinhua Street, and newly developed mixed-use areas in the eastern districts. The variety of urban contexts within the study area provides rich samples for analyzing spatial variations in street vitality.

The street network in the study area exhibits a hierarchical structure, comprising 1643 street segments with a total length of 680.3 km. This network includes traditional organic patterns in the historical core, grid-like layouts in planned districts, and contemporary street designs in newly developed areas. The diversity in street patterns and urban functions makes Hohhot an excellent laboratory for investigating how built environment perception influences street vitality across different spatial and functional contexts. The study area’s climate and cultural characteristics also contribute to its significance for street vitality research. With a semi-arid continental climate (annual average temperature 6.7 °C, annual precipitation 400 mm), the area experiences distinct seasonal variations that affect street life patterns. Furthermore, as a city where traditional Mongolian culture meets modern urban development, Hohhot offers unique insights into how cultural heritage and contemporary urban functions interact to shape street vitality.

2.2. Data Sources and Processing

2.2.1. Road Network Data

The street network data was obtained from OpenStreetMap (“https://www.openstreetmap.org (accessed on 13 May 2023)”) for 2019, 2021, and 2023. Given the complexity of the raw data and the presence of substantial irrelevant information, we implemented a systematic data cleaning procedure using ArcGIS 10.8. This process involved network topology correction, ensuring road continuity, and segmentation at intersections to create analytically viable street segments. Using the road centerlines as reference points, we established buffer zones according to road hierarchy: 55 m for expressways and arterial roads, 50 m for sub-arterial roads, and 30 m for local streets [39]. These buffer zones encompassed the three-dimensional street spaces defined by building enclosures, as well as adjacent commercial establishments and service facilities that potentially influence street vitality.

The resulting dataset comprises 1643 street segments with a total length of 526.8 km, including detailed attributes such as street hierarchies, intersection densities, and network connectivity information. Quality control measures included manual verification against local planning documents from the Hohhot Urban Planning Bureau to ensure both spatial accuracy and attribute consistency of the network data. Additionally, we performed cross-temporal validation to maintain consistency in street segment definitions across the three time points, enabling reliable longitudinal analysis.

2.2.2. Mobile Phone Signaling Data

Mobile phone signaling data was provided by Hohhot’s mobile telecommunications operator, capturing user–network interactions that precisely reflect urban population spatial distribution and mobility patterns [40]. Recognizing the influence of seasonal and weather conditions on human activity, we collected data for typical workdays and weekends in June of 2019, 2021, and 2023, with selected days having similar temperature ranges (12–29 °C) to ensure environmental comparability. The dataset encompasses records from 4378 base stations throughout the study area, containing only timestamps and geolocations of mobile communication behaviors, ensuring data anonymity and nonsensitivity. Our processing workflow involved hourly aggregation of signaling data, construction of Thiessen polygons around base stations to define service areas, and spatial overlay analysis with the road network to determine the number of people within each street buffer zone.

Data quality control procedures included: (1) systematic filtering of invalid records resulting from signal drift and ping-pong effects between adjacent base stations, (2) temporal normalization to account for daily activity cycles, (3) spatial interpolation to address areas with sparse base station coverage, and (4) population calibration using official census data to correct potential sampling biases. We then normalized activity density by street length and calculated temporal activity patterns for each segment to ensure comparable measurements across the study area and time periods.

2.2.3. POI Data

Point of Interest (POI) data, which effectively represents geographic entities in the real world and offers advantages in describing site spatial distributions, has been extensively applied in built environment assessments [41]. For this study, we collected POI data for 2019, 2021, and 2023 from Gaode Maps API (“https://lbs.amap.com/ (accessed on 2 April 2023)”), China’s largest online map service provider. Following preliminary data cleaning to remove duplicates and correct locational errors, we classified POIs according to the “The Urban Land Classification and Construction Land Planning Standards” (GB50137-2011) and Hohhot’s specific conditions. Eight major categories were established: residential, commercial facilities, financial and insurance, administrative offices, scientific–educational–cultural, medical and sports, transportation facilities, and green spaces/squares. The final dataset included 53,754 points for 2019, 54,776 points for 2021, and 54,828 points for 2023, all filtered based on the road network buffer zones.

Quality assurance procedures included: (1) geocoding verification to ensure accurate spatial positioning, (2) categorical standardization to maintain consistent classification across years, (3) density normalization to compensate for varying point densities across urban areas, and (4) cross-validation with satellite imagery and field surveys for a sample of locations to verify POI existence and accuracy. POI clustering analysis was performed to identify functional zones, and POI diversity indices were calculated to assess mixed-use characteristics at the street segment level.

2.2.4. Street View Imagery

Street view images were systematically sourced from Baidu Maps Open Platform (“https://lbsyun.baidu.com/ (accessed on 28 October 2023)”). Baidu Street View provides 360-degree panoramic images from an approximate “human perspective,” presenting rich and detailed urban street scenes that can effectively substitute human perception on streets [42,43]. Due to data update constraints, we accessed the server interface separately in 2019, 2021, and 2023 to obtain street view images for Hohhot’s central urban area. Our sampling strategy established points at 50 m intervals along all street segments. Using PyCharm Community Edition (Version 2022.2.3) programming to simulate human visual fields with standardized parameters, we collected street view images from four cardinal directions (front, back, left, right) at each sampling point. After rigorous data cleaning to remove poor-quality images (blurred, overexposed, or obstructed views), we obtained 98,493 valid street view images across the three time periods.

To accurately reflect the holistic perception of urban streetscapes, images from all four directions at each sampling point were stitched to create panoramic views. We then employed the Mask2Former model pre-trained on the MIT ADE20K dataset to perform semantic segmentation on these panoramic images, identifying and measuring the proportions of street space elements. Mask2Former, proposed by Cheng et al. [44] in 2021, incorporates Transformer architecture and mask classification mechanisms to extract multi-scale image features, effectively reducing misclassification cases and representing one of the leading algorithms in image recognition in terms of classification accuracy.

The image processing workflow included: (1) color calibration and lighting normalization to compensate for varying photography conditions, (2) distortion correction to address wide-angle lens effects, (3) seasonal standardization to minimize vegetation appearance variations across different months, and (4) confidence thresholding to ensure only high-confidence segmentation results were included in subsequent analyses. The resulting quantitative measurements of streetscape elements were validated through manual verification of a random 5% sample of images, achieving over 90% classification accuracy. An example of a downloaded street view image is shown in Figure 3.

All these data sources were systematically integrated into a unified spatio-temporal database using street segments as the basic analytical unit. This integration enables comprehensive longitudinal analysis of the relationships between built environment perceptions and street vitality while maintaining spatial and temporal consistency across different data types. Cross-dataset validation was performed to ensure alignment between mobile activity patterns, POI distributions, and streetscape characteristics, providing a robust foundation for analyzing the temporal evolution of perception–vitality relationships.

2.3. Methodology

This study adopts an innovative multi-level interpretative framework that combines Multi-scale Geographically Weighted Regression (MGWR) with XGBoost-SHAP analysis to comprehensively examine the spatial heterogeneity and nonlinear characteristics of perception–vitality relationships. This methodological integration represents a novel contribution to urban analytics, addressing complementary analytical needs that cannot be satisfied by either approach independently. The rationale for selecting this specific combination over alternative approaches lies in addressing fundamental limitations of conventional urban analysis methods. Previous urban studies have typically employed spatial regression methods (such as GWR) or machine learning techniques (such as random forest or SVM) separately [36,37]. However, traditional global regression models assume spatial stationarity and linear relationships, failing to capture the complex reality of urban environments where identical interventions may generate vastly different outcomes across space and intensity levels [38]. While local spatial methods like MGWR can address spatial heterogeneity, they inherently assume locally linear relationships and cannot capture threshold effects or interaction dynamics. Conversely, machine learning approaches excel at modeling nonlinear relationships but typically lack spatial interpretation capabilities, providing limited insights for spatially targeted urban interventions [29].

Our MGWR-SHAP framework overcomes these individual limitations by providing complementary analytical perspectives: MGWR quantifies the spatial variation of the perception–vitality relationship across urban space and its evolution over time. By employing adaptive bandwidth selection to explore scale dependencies, it captures the spatially nonstationary relationships operating at different geographic scales, revealing whether perception dimensions operate at local neighborhood scales or broader urban district levels [45,46]. Meanwhile, XGBoost-SHAP identified nonlinear threshold effects, feature interactions, and complex response patterns, while preserving the interpretability of the analytical results. This is crucial for understanding optimal intervention intensity and avoiding adverse effects, thereby providing essential insights for formulating optimal intervention strategies. The SHAP interpretation framework provides feature importance rankings and interaction visualizations that complement MGWR’s spatial coefficient maps, enabling both locally targeted and optimally calibrated urban design strategies [47,48]. This combination enables simultaneous examination of “where” relationships vary (spatial heterogeneity) and “how” they vary (nonlinear patterns), providing a more complete understanding of complex urban phenomena than single-method approaches.

2.3.1. Street Vitality and Built Environment Perception Measurement

We developed a comprehensive measurement framework incorporating multiple dimensions of both street vitality and built environment perceptions, as detailed in Table 1. Street vitality is conceptualized as a multi-dimensional construct encompassing social, economic, and cultural aspects [49,50]. Our selection of street vitality indicators is guided by both theoretical foundations and practical measurement considerations. Following Jacobs [51] and contemporary urban vitality research [52,53,54,55], we selected four complementary indicators (Population Density Index, Vitality Stability Index, Commercial Facility Density, and Cultural Facility Density) that collectively capture the presence of people, temporal consistency of activities, economic functions, and cultural significance of urban streets. These dimensions interact to create the complex phenomenon of urban street vitality, with each representing a distinct aspect of how streets function as social spaces.

Supplementary Materials Table S1 provides descriptive statistics of the composite vitality index across the three years (2019–2023), while Table S2 presents statistics for each of the four individual indicators. Supplementary Materials Figures S1 and S2 visualize the temporal evolution of these measurements through box plots and statistical distributions, enabling comprehensive examination of vitality patterns across time.

For built environment perception, our framework draws from environmental psychology and urban design theory, recognizing that pedestrian spatial perception significantly influences behavior in urban settings [15]. Based on both physiological and psychological human needs in street environments, we identified four key perception dimensions: comfort, safety, convenience, and pleasure. These perception dimensions represent the multi-faceted ways in which people experience and respond to built environments [31,56,57], with each dimension activating different psychological and behavioral responses. Recognizing that environmental cognition emerges from subject–object interaction and cannot be fully captured by objective measures alone [58], we incorporated both objective and subjective indicators in our measurement approach.

Based on relevant research [26,59,60,61,62,63,64,65,66,67], we selected 17 quantifiable indicators that comprehensively describe built environment perception, including 13 objective indicators reflecting physical street composition and 4 subjective indicators focusing on pedestrian experience. Table 1 details these indicators and their calculation methods. The entropy weight method was applied to assign appropriate weights within each perception dimension, creating composite indices that capture the complex nature of how people perceive urban environments. We conducted a detailed Variance Inflation Factor (VIF) analysis of the four perceptual dimensions for 2019, 2021, and 2023. All VIF values at all time points are below 5 (see Supplementary Materials Table S3), indicating that each dimension remains sufficiently independent despite some correlation between the dimensions.

Table 1. Calculation methods for street vitality and built environment perception variables.

Category	Variable	Sub-Variable	Formula	Description
Street Vitality	Social Vitality	Population Density Index (PDI)	$S{V}_n={\displaystyle \frac{{\sum }_{i=1}^n{m}_i}{S}}$	$m_i$: number of users at time $i$; $S$: street area
		Vitality Stability Index (VSI) [52,53]	$\mathrm{S}{V}_{std}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{\left(v_i-\overline{v}\right)}^2}{n}}}$	$v_i$: vitality at time $i$; $n$: number of time periods
	Economic Vitality	Commercial Facility Density (CFD)	$E{V}_n={\displaystyle \frac{{\sum }_{i=1}^n{e}_i}{L}}$	$e_i$: number of commercial facilities; $L$: street length
	Cultural Vitality	Cultural Facility Density (CuFD) [54,55]	$C{V}_n={\displaystyle \frac{{\sum }_{i=1}^n{c}_i}{L}}$	$c_i$: different types of cultural facilities
Built Environment Perception	Comfort Perception	Green View Index (GVI)	$GV{I}_n={\displaystyle \frac{G_n}{A_n}}$	$G_n$: green pixels; $A_n$: total pixels
		Sky View Index (SVI) [26]	$SV{I}_n={\displaystyle \frac{V_n}{A_n}}$	$V_n$: sky pixels
		Interface Enclosure Index (IEI) [59]	$IE{I}_n=-{\displaystyle \frac{O_n}{A_n}}$	$O_n$: building/wall pixels
		Street Cleanliness Index (SCI) [60]	Machine Learning Score	Based on street view image analysis
	Safety Perception	Sidewalk Visibility Index (SwVI) [61]	$SwV{I}_n={\displaystyle \frac{S_n}{A_n}}$	$S_n$: sidewalk pixels
		Traffic Safety Facility Index (TSFI) [62]	$TSF{I}_n={\displaystyle \frac{T_n}{A_n}}$	$T_n$: safety facility pixels
		Vehicle Impact Index (VII) [61]	$VI{I}_n={\displaystyle \frac{C_n}{A_n}}$	$C_n$: vehicle pixels
		Pedestrian Safety Index (PSI) [63]	Machine Learning Score	Based on street view image analysis
Built Environment Perception	Convenience Perception	Function Density Index (FDI)	$FD{I}_n={\displaystyle \frac{PO{I}_{num}}{L}}$	$PO{I}_{num}$: number of POIs
		Land Use Diversity Index (LDI) [64]	$LD{I}_n=\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \sum _{\mathrm{q}=1}^{\mathrm{p}}}{\mathrm{P}}_{\mathrm{i}\mathrm{q}}\mathrm{l}\mathrm{n}{\mathrm{P}}_{\mathrm{i}\mathrm{q}}\right)$	$p_i$: number of type $i$ POIs
		Public Transit Index (PTI) [61]	$PT{I}_n={\displaystyle \frac{PO{I}_{Public}}{L}}$	$PO{I}_{Public}$: number of transit POIs
		Facility Convenience Index (FCI) [65]	$FC{I}_n={\displaystyle \frac{p_n}{A_n}}$	$p_n$: facility pixels
	Pleasure Perception	Color Richness Index (CRI) [66]	$CR{I}_n=\left|\left({\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\left({\displaystyle \frac{S_i}{S_{total}}}\right)}^2\right)-1\right|$	$S_i$: area of color $i$; $S_{total}$: total analyzed area
		Feature Business Index (FBI) [65]	$FB{I}_n={\displaystyle \frac{PO{I}_{feature}}{L}}$	$PO{I}_{feature}$: number of featured businesses
		Landscape Aesthetic Index (LAI) [63]	Machine Learning Score	Based on street view image analysis
		Cultural Atmosphere Index (CAI) [67]	Machine Learning Score	Based on street view image analysis

Note: Vitality metrics use POIs as direct activity generators, while perception metrics use POIs as environmental elements that shape human perceptual experiences. And all machine learning scores are derived from trained perception models using street view imagery analysis.

Table 1. Calculation methods for street vitality and built environment perception variables.

Category	Variable	Sub-Variable	Formula	Description
Street Vitality	Social Vitality	Population Density Index (PDI)	$S{V}_n={\displaystyle \frac{{\sum }_{i=1}^n{m}_i}{S}}$	$m_i$: number of users at time $i$; $S$: street area
		Vitality Stability Index (VSI) [52,53]	$\mathrm{S}{V}_{std}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{\left(v_i-\overline{v}\right)}^2}{n}}}$	$v_i$: vitality at time $i$; $n$: number of time periods
	Economic Vitality	Commercial Facility Density (CFD)	$E{V}_n={\displaystyle \frac{{\sum }_{i=1}^n{e}_i}{L}}$	$e_i$: number of commercial facilities; $L$: street length
	Cultural Vitality	Cultural Facility Density (CuFD) [54,55]	$C{V}_n={\displaystyle \frac{{\sum }_{i=1}^n{c}_i}{L}}$	$c_i$: different types of cultural facilities
Built Environment Perception	Comfort Perception	Green View Index (GVI)	$GV{I}_n={\displaystyle \frac{G_n}{A_n}}$	$G_n$: green pixels; $A_n$: total pixels
		Sky View Index (SVI) [26]	$SV{I}_n={\displaystyle \frac{V_n}{A_n}}$	$V_n$: sky pixels
		Interface Enclosure Index (IEI) [59]	$IE{I}_n=-{\displaystyle \frac{O_n}{A_n}}$	$O_n$: building/wall pixels
		Street Cleanliness Index (SCI) [60]	Machine Learning Score	Based on street view image analysis
	Safety Perception	Sidewalk Visibility Index (SwVI) [61]	$SwV{I}_n={\displaystyle \frac{S_n}{A_n}}$	$S_n$: sidewalk pixels
		Traffic Safety Facility Index (TSFI) [62]	$TSF{I}_n={\displaystyle \frac{T_n}{A_n}}$	$T_n$: safety facility pixels
		Vehicle Impact Index (VII) [61]	$VI{I}_n={\displaystyle \frac{C_n}{A_n}}$	$C_n$: vehicle pixels
		Pedestrian Safety Index (PSI) [63]	Machine Learning Score	Based on street view image analysis
Built Environment Perception	Convenience Perception	Function Density Index (FDI)	$FD{I}_n={\displaystyle \frac{PO{I}_{num}}{L}}$	$PO{I}_{num}$: number of POIs
		Land Use Diversity Index (LDI) [64]	$LD{I}_n=\mathrm{e}\mathrm{x}\mathrm{p}\left(-{\displaystyle \sum _{\mathrm{q}=1}^{\mathrm{p}}}{\mathrm{P}}_{\mathrm{i}\mathrm{q}}\mathrm{l}\mathrm{n}{\mathrm{P}}_{\mathrm{i}\mathrm{q}}\right)$	$p_i$: number of type $i$ POIs
		Public Transit Index (PTI) [61]	$PT{I}_n={\displaystyle \frac{PO{I}_{Public}}{L}}$	$PO{I}_{Public}$: number of transit POIs
		Facility Convenience Index (FCI) [65]	$FC{I}_n={\displaystyle \frac{p_n}{A_n}}$	$p_n$: facility pixels
	Pleasure Perception	Color Richness Index (CRI) [66]	$CR{I}_n=\left|\left({\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\left({\displaystyle \frac{S_i}{S_{total}}}\right)}^2\right)-1\right|$	$S_i$: area of color $i$; $S_{total}$: total analyzed area
		Feature Business Index (FBI) [65]	$FB{I}_n={\displaystyle \frac{PO{I}_{feature}}{L}}$	$PO{I}_{feature}$: number of featured businesses
		Landscape Aesthetic Index (LAI) [63]	Machine Learning Score	Based on street view image analysis
		Cultural Atmosphere Index (CAI) [67]	Machine Learning Score	Based on street view image analysis

Note: Vitality metrics use POIs as direct activity generators, while perception metrics use POIs as environmental elements that shape human perceptual experiences. And all machine learning scores are derived from trained perception models using street view imagery analysis.

For each street segment $i$, we calculate a composite Vitality Index ($VI$) using the entropy method [68,69]:

$$V{I}_i={\displaystyle \sum _{j=1}^n} \left(w_j\times v_{ij}\right)$$

(1)

where $w_j$ represents the entropy-derived weight for dimension $j$, and $v_{ij}$ is the normalized score of dimension $j$ for street segment $i$.

Built environment perception is measured through a systematic integration of objective spatial metrics and computer vision analysis of street view imagery. For each perception dimension, we construct composite indices ($PI$) that capture both physical characteristics and human-scale perceptual qualities:

$$P{I}_i={\displaystyle \sum _{k=1}^K} \left({\alpha }_k\times p_{ik}\right)$$

(2)

where $P{I}_i$ represents the perception index for dimension $i$, ${\alpha }_k$ are weights derived from factor analysis, and $p_{ik}$ are the standardized scores of individual perception metrics.

2.3.2. MGWR Analysis

To capture spatial heterogeneity in perception–vitality relationships, we employ MGWR with an adaptive bandwidth selection approach [37]. The model is specified as:

$$y_i={\beta }_0\left(u_i,v_i\right)+\sum _{k=1}^k{\beta }_k\left(u_i,v_i\right)x_{ik}+{ϵ}_i$$

(3)

where $\left(u_i,v_i\right)$ represents the spatial coordinates of location $i$, ${\beta }_k\left(u_i,v_i\right)$ are locally varying coefficients, and $x_{ik}$ are the built environment perception measures.

The optimal bandwidth for each variable is determined through cross-validation:

$$CV={\displaystyle \sum _{i=1}^n} [y_i-{\widehat{y}}_{\ne i}(b){]}^2$$

(4)

where ${\widehat{y}}_{\ne i}\left(b\right)$ is the fitted value for observation $i$ using bandwidth $b$ with observation $i$ omitted from the calibration.

2.3.3. XGBoost-SHAP Framework

To explore nonlinear relationships and complex interactions, we implement an XGBoost model, a gradient boosting framework that excels at capturing nonlinear patterns in data [70,71,72]. The model is configured with carefully tuned hyperparameters to balance model complexity and generalization ability, including the number of trees, learning rate, and tree depth. This configuration enables the model to capture complex nonlinear relationships while avoiding overfitting.

The XGBoost model is complemented by SHapley Additive exPlanations (SHAP) analysis to interpret feature importance and interaction effects. SHAP values are calculated as:

$${\varphi }_j=\sum _{S\subseteq F\setminus j} {\displaystyle \frac{\left|S\right|!\left(\left|F\right|-\left|S\right|-1\right)!}{\left|F\right|!}}\left[v\left(S\cup j\right)-v\left(S\right)\right]$$

(5)

where ${\varphi }_j$ is the SHAP value for feature $j$, $S$ represents all possible feature combinations, and $v\left(S\right)$ is the prediction for feature set $S$.

This integrated analytical framework combines multiple methodological strengths to provide comprehensive insights into the perception–vitality relationship. Through MGWR coefficients, we quantify the varying spatial relationships at different scales, while XGBoost predictions reveal complex nonlinear patterns that might be missed by traditional linear approaches. The SHAP analysis further enriches our understanding by illuminating feature interactions and their spatial manifestations. Together, these analytical components enable both detailed local analysis and broader pattern recognition, supporting the development of nuanced, context-sensitive urban design strategies [47,48].

The combination of these methods provides a robust approach for understanding the complex relationships between built environment perceptions and street vitality, while maintaining interpretability for practical applications in urban design and planning. The framework’s multi-method nature allows for cross-validation of findings and provides complementary insights into both spatial and nonlinear aspects of perception–vitality relationships, ultimately supporting more informed and effective urban design decisions. To evaluate the superiority of the XGBoost model, we conducted comprehensive comparisons with alternative machine learning algorithms, including Gradient Boosting Decision Tree (GBDT), random forest, and Support Vector Machine (SVM). As detailed in Supplementary Materials Table S5, XGBoost consistently outperformed these alternatives across multiple evaluation metrics, demonstrating superior predictive accuracy and more effective capture of the complex nonlinear relationships in our spatio-temporal dataset.

3. Results

This section presents our empirical findings through a three-stage analysis framework that examines both spatial heterogeneity and temporal evolution of perception–vitality relationships from 2019 to 2023. First, we analyze the spatial distribution patterns of street vitality and built environment perceptions across different land use types, revealing a pronounced center–periphery gradient and significant temporal increases in vitality (21.4% overall growth). Second, we employ MGWR analysis to investigate the evolving spatial heterogeneity in perception–vitality relationships, demonstrating a trend toward spatial homogenization (bandwidth convergence from varied 365–1537 m to uniform 1642 m) and revealing dramatic coefficient sign reversals, particularly for safety perception (from −0.050 in 2019 to +0.069 in 2023). Third, we utilize machine learning techniques to uncover complex nonlinear interactions, with XGBoost models achieving superior performance (average R² = 0.444) compared to GBDT, random forest, and SVM alternatives (see Supplementary Materials Table S5). The SHAP analysis further reveals threshold effects and inverse U-shaped relationships that would remain undetected in conventional approaches, particularly for convenience perception which demonstrates a complex curvilinear relationship with vitality. Together, these analyses demonstrate how perception–vitality relationships have evolved from highly differentiated patterns in 2019 toward greater stability and homogeneity by 2023, suggesting a maturing urban environment with increasingly standardized environmental effects.

3.1. Temporal Evolution and Spatial Distribution Patterns Across Land Use Types

Using the entropy weight method, we constructed comprehensive indices for street vitality across multiple years (2019, 2021, and 2023) to analyze both temporal evolution and spatial distribution patterns. The analysis of spatial distribution maps (Figure 4) provides insights into how street vitality varies across different urban locations and land use types over time.

Street vitality exhibits a pronounced center–periphery gradient pattern across the study area. The highest vitality values are consistently concentrated in the central business districts, particularly around the Zhongshan Road Commercial District and historical cultural areas such as the Dazhao Temple District. These areas benefit from the convergence of commercial activities, cultural facilities, and pedestrian flows. Vitality values gradually decrease from these central nodes toward peripheral areas, with the lowest values observed in outer residential and industrial zones, reflecting the transition from mixed-use urban cores to more functionally specialized areas.

The temporal analysis from 2019 to 2023 reveals a steady increase in overall street vitality in Hohhot’s central urban area. Descriptive statistics show that the mean vitality value increased from 0.083 in 2019 to 0.087 in 2021 and reached 0.101 in 2023, representing an approximate 21.4% growth. This upward trend indicates consistent improvement in urban street environments and public space vitality in recent years.

Notably, the spatial extent of high-vitality areas (orange and red regions on the maps) has expanded over the years, particularly in the northward and eastward directions from the city center. This suggests that vitality growth is not spatially homogeneous but follows specific expansion patterns tied to urban development priorities.

The land use analysis (Figure 4d) classifies the study area into five primary types: commercial land, industrial land, public management and service land, residential land, and transportation land. When examining street vitality across these different land use types (Figure 4e), significant variations emerge in their performance and temporal evolution. Commercial land consistently exhibits the highest street vitality values, with three-year means of 0.100 (2019), 0.103 (2021), and 0.119 (2023). This superior performance can be attributed to the high pedestrian density, diverse facilities, and rich functional mix that characterize commercial areas, creating favorable conditions for human activities and interactions. In contrast, industrial land shows the lowest street vitality across all three years, with means of 0.071, 0.075, and 0.084, respectively. This relatively lower performance aligns with the typically monofunctional nature of industrial zones, where reduced pedestrian flows, limited diversity of uses, and relatively lower public space quality often constrain vibrant street life. Between these two extremes, public management and service land, residential land, and transportation land display intermediate street vitality levels, suggesting these functional zones play similar supportive roles in urban activity patterns despite their different primary functions.

The temporal evolution of vitality across land use types reveals another interesting pattern. All five land use categories experienced increased vitality from 2019 to 2023, though with varying magnitudes of improvement. Commercial land demonstrated the most substantial absolute growth in vitality (19.1% increase), while transportation land showed the highest relative improvement (27.4% increase). These differential growth rates likely reflect the varying impacts of urban renewal initiatives, commercial development, and infrastructure improvements across different functional zones. The consistent improvement across all land use types, regardless of their baseline levels, suggests a holistic enhancement of the urban environment throughout the study period.

The error bars in Figure 4e further reveal considerable variability within each land use category, particularly pronounced in commercial and residential lands. This internal heterogeneity suggests that, even within the same land use classification, street vitality can vary significantly based on micro-scale environmental characteristics, design qualities, and local context factors. This finding highlights the importance of considering both macro-level land use policies and micro-level urban design interventions when seeking to enhance street vitality.

These spatial distribution patterns and land use analyses collectively demonstrate the complex nature of street vitality and its close association with urban structure. Central areas typically perform strongly across multiple dimensions, while peripheral areas exhibit more diverse patterns, providing direction for environmental enhancement in different functional zones. The overall increase in street vitality over time indicates a positive trend in urban environmental quality improvement, though the disparities across different areas and land use types also highlight priority zones for future urban planning attention.

The spatial analysis of built environment perceptions in 2023 reveals complex and varied patterns across the four dimensions examined (Figure 5) Comfort perception demonstrates higher values in the eastern and southeastern regions of the city, where better green coverage and environmental maintenance contribute to a more pleasant urban experience. However, the central commercial districts, despite their high vitality, show relatively lower comfort levels due to increased building density and interface enclosure.

Safety perception manifests a centralized pattern with peak values in the core urban area, characterized by well-maintained infrastructure and pedestrian facilities. This pattern is interrupted along major traffic arteries in the southern and peripheral areas, where higher vehicular traffic density creates perceived safety challenges. The convenience perception displays a strong centralized distribution, with highest values in areas well-served by public facilities and services, gradually diminishing toward the urban periphery.

Pleasure perception shows a distinctive pattern closely aligned with the city’s cultural and historical assets. The highest pleasure perception values are found in the central area, particularly around historical districts and cultural landmarks. This distribution reflects the rich cultural heritage and aesthetic quality of these areas, while the peripheral zones, dominated by residential and industrial functions, exhibit lower pleasure perception values, indicating potential areas for future urban design interventions.

The integration of land use patterns (Figure 4d) with perception dimension distributions (Figure 5) reveals important relationships between urban function and environmental perception. When overlaying these spatial distributions, we observe that commercial land areas generally correspond with high convenience perception and pleasure perception values, particularly in the central urban districts, though they often exhibit relatively lower comfort perception scores. This suggests that, while commercial zones effectively support functional accessibility and cultural experience, they may face challenges in providing comfortable environments due to higher density development and anthropogenic pressures. Residential lands show more balanced perception profiles, with moderate to high comfort perception in eastern residential zones where newer development incorporates more green infrastructure and open spaces. Industrial lands consistently demonstrate lower values across all perception dimensions, most notably in comfort and pleasure perceptions, indicating comprehensive environmental quality issues in these areas that extend beyond their low vitality scores. Public management and service lands exhibit particularly strong performance in safety perception, reflecting the typically well-maintained infrastructure and regulated environments of institutional areas. Transportation lands show significant internal variation in perception profiles, with major corridors scoring high on convenience but lower on comfort and safety, while secondary transportation areas demonstrate more balanced perception characteristics. These patterns indicate that perception dimensions have varying relationships with different land use types, suggesting that environmental enhancement strategies should be tailored to address the specific perception deficiencies associated with each urban function.

3.2. MGWR Analysis Results: Temporal Evolution of Spatial Heterogeneity (2019–2023)

To address the potential spatial nonstationarity in the relationships between built environment perceptions and street vitality, we employed the MGWR model for three consecutive periods (2019, 2021, and 2023), which allows regression coefficients to vary across space and time. This longitudinal approach reveals not only significant spatial heterogeneity in how different perception dimensions influence street vitality across Hohhot’s central urban area but also captures the temporal evolution of these relationships, providing insights that would be masked by traditional global regression methods or cross-sectional analysis. The MGWR models demonstrated satisfactory explanatory power with R² values of 0.173 for 2019, 0.198 for 2021, and 0.180 for 2023, indicating consistent performance across the study period. Notably, all variables across all three models showed statistical significance (p < 0.05), confirming the reliable contribution of all perception dimensions to explaining street vitality variations.

3.2.1. Temporal Shifts in Bandwidth and Model Structure

The MGWR model results in

Table 2. Summary of Multi-scale Geographically Weighted Regression results from 2019 to 2023.

Year	Variable	Bandwidth	ENP_j	Adj t-val (95%)	Adj Alpha (95%)	Mean	STD	Min	Median	Max
2019	Constant	178.000	24.802	3.093	0.002 **	0.111	0.288	−0.533	0.198	0.639
	Comfort Perception	565.000	7.956	2.736	0.006 **	0.006	0.053	−0.111	0.001	0.119
	Safety Perception	1096.000	3.190	2.419	0.016 *	−0.050	0.045	−0.13	−0.053	0.064
	Convenience Perception	386.000	8.459	2.756	0.006 **	0.054	0.141	−0.138	0.011	0.377
	Pleasure Perception	1537.000	1.549	2.143	0.032 *	0.032	0.012	0.011	0.031	0.065
2021	Constant	79.000	60.906	3.352	0.001 ***	0.003	0.326	−0.683	0.047	0.802
	Comfort Perception	1477.000	2.219	2.283	0.023 *	0.003	0.032	−0.045	0.000	0.053
	Safety Perception	1642.000	1.409	2.105	0.035 *	0.073	0.007	0.060	0.073	0.085
	Convenience Perception	1642.000	1.145	2.019	0.044 *	−0.054	0.003	−0.058	−0.054	−0.046
	Pleasure Perception	1619.000	1.408	2.104	0.036 *	0.086	0.009	0.072	0.084	0.103
2023	Constant	93.000	51.762	3.306	0.001 ***	0.006	0.301	−0.637	0.016	0.668
	Comfort Perception	1642.000	1.534	2.139	0.033 *	0.043	0.002	0.037	0.043	0.047
	Safety Perception	1642.000	1.421	2.108	0.035 *	0.069	0.007	0.054	0.070	0.079
	Convenience Perception	1642.000	1.151	2.021	0.043 *	0.058	0.003	0.054	0.057	0.068
	Pleasure Perception	1642.000	1.285	2.067	0.039 *	0.033	0.003	0.031	0.032	0.044

Note: ENP_j represents the effective number of parameters; STD is the standard deviation of the local coefficient estimates. Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05.

reveal significant temporal shifts in both the spatial scale and magnitude of perception–vitality relationships. One of the most striking temporal patterns is the evolution of bandwidths across the three time periods. In 2019, perception dimensions operate at distinctly different spatial scales, with comfort perception (565 m) and convenience perception (386 m) showing highly localized effects, while pleasure perception (1537 m) exhibits a much broader influence. By 2021, a convergence trend emerges with comfort perception bandwidth expanding to 1477 m, and by 2023, all perception dimensions operate at virtually identical spatial scales (1642 m), suggesting a homogenization of their spatial influence over time.

The effective number of parameters (ENP_j) shows a corresponding temporal pattern, decreasing across all variables from 2019 to 2023. This indicates that, while the spatial heterogeneity of perception–vitality relationships was highly pronounced in 2019, it gradually diminished over the study period, pointing to a possible standardization of urban environments or convergence in how perceptions influence vitality across the city. This evolution possibly reflects the impact of urban renewal and standardized development practices implemented throughout the city during this period.

3.2.2. Temporal Evolution of Coefficient Magnitudes and Signs

The mean coefficients for each perception dimension demonstrate significant temporal shifts in their influence on street vitality. Safety perception underwent the most dramatic transformation, shifting from a negative influence in 2019 (mean = −0.050) to a strong positive influence in 2021 (mean = 0.073) and 2023 (mean = 0.069). This reversal suggests fundamental changes in how safety factors impact urban vitality, possibly reflecting improved safety infrastructure or changes in public perception of safety over time.

Convenience perception displays an equally remarkable evolution, shifting from a positive influence in 2019 (mean = 0.054) to a negative influence in 2021 (mean = −0.054), then returning to a positive influence in 2023 (mean = 0.058). This nonlinear temporal pattern indicates a complex relationship that may reflect initial oversaturation of convenience facilities by 2021, followed by better integration into the urban fabric by 2023.

Comfort perception shows a more subtle but important evolution, from a slightly positive influence in 2019 (mean = 0.006) to near zero in 2021 (mean = 0.003), before strengthening significantly in 2023 (mean = 0.043). This progressive strengthening suggests increasing public preference for comfortable environments as basic urban needs are satisfied, an evolution consistent with Maslow’s hierarchy applied to urban settings.

Pleasure perception demonstrates a particularly interesting pattern, with influence peaking in 2021 (mean = 0.086) before declining in 2023 (mean = 0.033), though remaining positive throughout. This pattern may reflect the temporary heightened importance of pleasurable environments during the pandemic period of 2021, followed by a rebalancing of priorities by 2023.

3.2.3. Spatial Patterns and Their Temporal Evolution

The spatial distribution maps of MGWR coefficients (Figure 6) reveal how the influence of each perception dimension varies across space and time, providing deeper insights into the evolving urban dynamics of Hohhot.

The constant term maps show persistent “hotspots” of unexplained vitality in the central urban areas across all three years, though the spatial configuration evolves. In 2019, high values concentrate in a compact central core, while 2021 shows a more dispersed pattern with stronger east–west differentiation. By 2023, the pattern resolidifies with concentrated high values in the commercial center but with new secondary centers emerging in northern areas, reflecting the spreading of urban core functions to new development zones over time.

The comfort perception coefficients display one of the most dramatic spatial transformations. In 2019, the map reveals a highly heterogeneous pattern with sharp local variations, positive influences concentrated in southeastern areas and negative influences in the urban core. By 2021, this pattern smoothens considerably with narrower coefficient ranges, though maintaining an east–west gradient. The 2023 map reveals a fundamental shift toward universal positive influence across the entire study area, with coefficients becoming remarkably homogeneous (0.037–0.047). This evolution suggests that comfort factors, initially having mixed and localized effects, have become consistently positive contributors to vitality throughout the city, reflecting possibly successful urban greening and environmental quality improvement initiatives.

Safety perception exhibits perhaps the most dramatic transformation in both coefficient signs and spatial patterns. The 2019 map shows predominantly negative coefficients (blue areas) across most of the study area, with only small pockets of positive influence in the south. By 2021, this pattern completely reverses, with universally positive coefficients and a clear spatial gradient from southeast (highest positive influence) to northwest. The 2023 map maintains this positive pattern but with a more nuanced spatial distribution, showing particularly strong influences in the eastern commercial and newly developed areas. This remarkable transformation suggests a fundamental shift in how safety factors contribute to vitality, possibly reflecting improvements in the city’s safety infrastructure and changing resident perceptions over this period.

Convenience perception displays a fascinating oscillation in its spatial patterns. The 2019 map shows a complex mosaic with strong positive influences in peripheral areas and negative influences in some central zones. By 2021, this pattern inverts dramatically to show universal negative influence with very little spatial variation, suggesting a consistent negative relationship between convenience facilities and vitality across the city. The 2023 map then shows another complete reversal, with universally positive coefficients and a new spatial concentration of stronger positive effects in the southern regions. This oscillating pattern potentially reflects a complex recalibration of convenience needs and provisions throughout the urban environment over time.

Pleasure perception displays a unique temporal evolution in its spatial pattern. The 2019 map shows a relatively homogeneous positive influence across the study area with subtle variations. In 2021, the coefficients strengthen substantially and develop a pronounced east–west gradient with strongest positive influences in the eastern commercial and cultural districts. By 2023, while remaining universally positive, the coefficients decrease in magnitude and develop a more nuanced spatial pattern with higher values in central–eastern areas corresponding to historical and cultural zones. This evolution suggests a growing but increasingly spatially differentiated importance of aesthetic and cultural factors in shaping street vitality.

Supplementary Materials Figure S3 provides a comprehensive statistical analysis of the evolution of the MGWR coefficients.

3.2.4. Synthesis of Temporal and Spatial Patterns

The comparative analysis across 2019, 2021, and 2023 reveals several important trends in how environmental perceptions influence street vitality. Our longitudinal analysis demonstrates a clear pattern of increasing spatial homogenization, as all perception dimensions show a consistent trend toward larger bandwidths and more homogeneous spatial patterns over time. This convergence suggests a gradual standardization of urban environmental conditions and their effects on vitality, possibly resulting from coordinated urban planning efforts and standardized development practices implemented across Hohhot during this period. This homogenization process may indicate successful urban integration, though it also raises questions about potential loss of neighborhood distinctiveness.

The temporal evolution further reveals remarkable sign reversals and nonlinear development in how perceptions influence vitality. Several dimensions, particularly safety and convenience perceptions, display complete reversals in their relationship with vitality over the study period. This finding fundamentally challenges static conceptions of environmental influences, indicating that the nature of perception–vitality relationships is dynamic and evolves with changing urban conditions and resident expectations. These nonlinear temporal patterns suggest complex adaptation processes as residents adjust their preferences and behaviors in response to evolving urban environments.

By 2023, we observe the emergence of more stable and consistent spatial distributions for all perception dimensions compared to the highly variable patterns of 2019. This stabilization suggests a maturing urban environment where perception–vitality relationships have achieved a form of equilibrium after a period of rapid change. The more consistent spatial patterns may reflect the completion of major urban development initiatives and the establishment of more stable activity patterns throughout the city, providing a clearer foundation for targeted future interventions.

The relative importance of different perception dimensions also shifts significantly over time, reflecting changing societal priorities. While pleasure perception dominated in 2021 with the strongest positive influence on vitality, comfort and safety perceptions gained substantial importance by 2023. This reprioritization may reflect broader societal shifts in environmental preferences, possibly influenced by pandemic experiences that temporarily elevated the importance of pleasurable environments during 2021, followed by a return to more fundamental comfort and safety concerns by 2023.

These temporal dynamics in spatial heterogeneity have important implications for urban planning and design. They demonstrate that perception–vitality relationships are not static properties of urban environments but dynamic phenomena that evolve with changing urban conditions, suggesting that environmental interventions need to be periodically reassessed and adjusted. The findings reinforce the importance of employing spatially and temporally sensitive approaches to urban design rather than applying uniform solutions across entire cities or maintaining them unchanged over time. Understanding these dynamic relationships enables more adaptive and responsive urban planning strategies that can anticipate and accommodate evolving resident needs and preferences.

3.3. Machine-Learning-Based Nonlinear Analysis

While the MGWR analysis revealed important spatial variations in perception–vitality relationships, it inherently assumes locally linear relationships between variables. To overcome this limitation and capture potential nonlinear effects and complex interactions, we employed a machine learning approach combining XGBoost with SHapley Additive exPlanations (SHAP) analysis. This complementary methodology not only achieved superior predictive performance but also provided deeper insights into the temporal evolution of nonlinear perception–vitality relationships across the three-year study period.

Our model comparison analysis demonstrated that XGBoost consistently outperformed alternative machine learning techniques across all three years (2019–2023). The XGBoost models achieved average R² values of 0.444 (with R² of 0.458 for 2019, 0.434 for 2021, and 0.441 for 2023), significantly exceeding the explanatory power of the MGWR models examined in Section 3.2 (see Supplementary Materials Table S4 for detailed performance metrics). This performance advantage was consistent across other metrics, with the XGBoost models maintaining lower Root Mean Square Error (RMSE mean: 0.056) and Mean Absolute Error (MAE mean: 0.042) values compared to GBDT, random forest, and SVM alternatives (see Supplementary Materials Table S5 for model comparison). This superior predictive capability suggests that nonlinear interactions between perception dimensions and street vitality are substantial, extending well beyond what traditional linear or locally linear approaches can capture.

The temporal comparison of SHAP values across the three years (see Supplementary Materials Figure S4 for detailed SHAP summary plots) reveals both consistency and evolution in how different perception dimensions influence street vitality nonlinearly. Pleasure perception consistently demonstrates the strongest positive impact on street vitality throughout the study period, though with interesting temporal fluctuations. In 2019, pleasure perception shows a balanced distribution of SHAP values with moderate magnitude; by 2021, its influence intensifies significantly and becomes more consistently positive, particularly for high feature values; then in 2023, while maintaining its position as the most influential dimension, its impact pattern becomes more nuanced with greater variability. This temporal pattern aligns with our MGWR findings in Section 3.2, where pleasure perception’s influence peaked in 2021 before moderating in 2023, suggesting a consistent phenomenon captured by both methodologies.

Safety perception displays perhaps the most dramatic temporal transformation in its nonlinear relationship with vitality. In 2019, it exhibits a predominantly negative influence with a distinctive bimodal distribution; by 2021, this pattern dramatically reverses to show strongly positive contributions, particularly at higher safety values; and in 2023, while maintaining positive influence, the relationship becomes more complex with greater spread in SHAP values. This remarkable evolution mirrors the coefficient sign reversals observed in our MGWR analysis, confirming through two different methodological approaches that safety perception underwent a fundamental transformation in how it influenced street vitality during this period.

Convenience perception reveals particularly complex nonlinear patterns that evolve over time. The SHAP summary in Supplementary Materials Figure S4 shows a distinctive polarized distribution with both positive and negative SHAP values in 2019; in 2021, this pattern shifts toward predominantly negative contributions; by 2023, the pattern reverses again to show largely positive influence with high feature values yielding particularly strong positive effects. This oscillating pattern aligns with the mean coefficient sign changes observed in our MGWR analysis, but the SHAP approach additionally reveals the underlying nonlinear complexity within each year, suggesting threshold effects and interaction dynamics not captured by the MGWR approach.

Comfort perception shows the most stable and consistent nonlinear relationship pattern across the three years, though with gradually increasing positive influence. In 2019, its SHAP values center closely around zero with slight positive tendency; by 2021, a more distinct positive pattern emerges for higher comfort values; and in 2023, its positive influence strengthens further while maintaining a similar nonlinear pattern. This progressive strengthening mirrors the temporal evolution of comfort perception coefficients in our MGWR analysis, suggesting a consistent trajectory toward greater importance of comfort factors in shaping street vitality.

The SHAP dependence plots (Figure 7) provide deeper insights into the specific forms of these nonlinear relationships across the three years. Comfort perception exhibits a particularly interesting nonlinear evolution, developing from a relatively flat relationship in 2019 to a more pronounced threshold effect by 2023, where its positive influence on vitality increases sharply after reaching moderate feature values. This suggests that basic comfort requirements must be met before significant vitality benefits can be realized, an insight impossible to derive from linear models.

Safety perception’s dependence plots dramatically illustrate its transformation, showing a predominantly negative relationship in 2019 that inverts to a positive J-shaped curve by 2021, with the strongest positive effects appearing at the highest safety values. By 2023, the relationship evolves further into a more complex S-shaped pattern, suggesting multiple transition points and sensitivity zones. This complex evolution indicates that safety perception’s influence on vitality operates through sophisticated threshold effects that vary temporally, requiring nuanced interpretation beyond simple linear correlations.

Convenience perception’s dependence plots reveal perhaps the most complex nonlinear patterns among all dimensions. In 2019, it displays an inverse U-shaped relationship where moderate convenience levels yield optimal vitality benefits; by 2021, this transforms into a predominantly negative relationship with diminishing returns; and in 2023, the pattern shifts again toward a positive exponential relationship where benefits accelerate at higher convenience levels. This finding suggests that convenience facilities’ relationship with vitality is highly contextual and potentially subject to saturation effects that evolve over time with changing urban conditions and resident expectations.

Pleasure perception’s dependence plots show a consistently positive but nonlinear relationship across all three years, with the steepest positive slope appearing in 2021. The 2019 plot reveals a threshold effect where benefits accelerate after moderate pleasure values; the 2021 plot shows a more consistent positive slope throughout the feature range; and the 2023 plot indicates a return to a more threshold-dependent pattern. This temporal cycle may reflect shifting priorities in how aesthetic and cultural elements contribute to vitality in different phases of urban development.

The spatial distribution of SHAP values (Figure 8) reveals how these nonlinear relationships manifest geographically across the urban landscape. Comfort perception’s spatial influence becomes increasingly homogeneous over time, with scattered positive hotspots in 2019 becoming more uniformly distributed by 2023. Safety perception shows the most dramatic spatial transformation, shifting from concentrated negative influence in central areas in 2019 to widely distributed positive influence by 2023. Convenience perception displays oscillating spatial patterns that reflect its changing relationship with vitality, while pleasure perception maintains relatively stable spatial patterns with consistent influence in historical and cultural districts.

These spatial distributions of SHAP values complement our MGWR coefficient maps from Section 3.2, but with important distinctions that highlight the added value of the machine learning approach. While MGWR identifies local coefficient variations assuming linear relationships within each locality, the SHAP maps reveal where nonlinear effects are most pronounced, capturing complex interaction effects that may not be evident in coefficient maps alone. This distinction is particularly evident for convenience perception, where the SHAP maps reveal nuanced center–periphery gradients that reflect the inverse U-shaped relationship identified in the dependence plots.

The integration of machine learning and spatial analysis provides several crucial insights beyond what either approach could offer independently. First, the temporal evolution of nonlinear relationships reveals adaptation processes in how residents respond to environmental perceptions, with relationships transforming from simple to complex patterns as urban environments mature. Second, the identification of threshold effects across multiple dimensions suggests that urban interventions may need to reach certain intensity levels before yielding meaningful vitality benefits. Third, the discovery of inverse U-shaped relationships, particularly for convenience perception, indicates potential diminishing returns or even negative consequences of excessive development in certain urban contexts.

These findings demonstrate the complementary nature of MGWR and XGBoost-SHAP approaches in understanding the complex dynamics of perception–vitality relationships. While MGWR effectively captures spatial heterogeneity and temporal evolution of linear effects, the machine learning approach reveals the underlying nonlinear patterns and interaction effects that contribute to these spatial variations. Together, they provide a more comprehensive understanding of how built environment perceptions influence street vitality across both space and time, emphasizing the need for urban interventions that are sensitive to both spatial context and nonlinear response thresholds.

4. Discussion

Our longitudinal analysis of Hohhot reveals how the evolving relationships between built environment perceptions and street vitality can inform more effective, responsive urban design strategies. Rather than treating these relationships as static, our findings demonstrate that successful urban interventions must adapt to the dynamic nature of how residents perceive and interact with their built environments. This section explores how these insights can address critical urban challenges facing rapidly developing cities.

4.1. Addressing Urban Form Optimization Through Spatial Heterogeneity Insights

The spatial heterogeneity in perception–vitality relationships provides crucial guidance for optimizing urban form in rapidly developing contexts. High-vitality central areas, characterized by strong pleasure and convenience perceptions, can support continued intensification and mixed-use development. Meanwhile, peripheral areas with lower vitality scores require targeted interventions that address specific perception deficiencies, including comfort improvements in eastern residential zones, safety enhancements along major traffic corridors, and convenience facility provision in underserved areas. This spatially differentiated approach enables more effective resource allocation and can prevent the creation of monofunctional peripheral areas that lack vibrancy.

4.2. Enhancing Public Space Design Through Nonlinear Threshold Understanding

The identification of nonlinear threshold effects fundamentally changes how we should approach public space interventions. The inverse U-shaped relationship discovered for convenience perception demonstrates that moderate provision of services and facilities generates optimal vitality benefits, while excessive provision can actually diminish street life. This finding has profound implications for commercial district planning, suggesting that planners should resist the temptation to maximize retail density and instead focus on achieving optimal facility mixes that support vibrant street environments without creating oversaturation.

Comfort perception’s threshold effects provide specific guidance for green infrastructure and public space improvements. Our analysis shows that comfort investments only begin generating significant vitality benefits after reaching moderate levels, suggesting that incremental improvements in poorly performing areas may be less effective than concentrated investments that achieve meaningful environmental transformations. For cities with limited resources, this finding supports focusing green infrastructure investments in areas where they can achieve threshold-level impacts rather than spreading improvements thinly across the urban area.

The S-shaped relationship observed for safety perception reveals the complex dynamics of security interventions in urban environments. The multiple transition points in this relationship suggest that cities can overcome initial safety deficits through comprehensive and sustained interventions in physical infrastructure and environmental design, but that excessive safety infrastructure can reduce street vitality by overcontrolling the environment.

4.3. Informing Socio-Spatial Equity Through Temporal Priority Understanding

The temporal evolution of perception priorities offers crucial insights for addressing socio-spatial equity in rapidly urbanizing cities. The shift from pleasure-dominated relationships (2021) to stronger comfort and safety influences (2023) suggests that residents’ environmental priorities evolve as cities mature, with basic needs eventually taking precedence over aesthetic considerations. This progression has important implications for equitable development strategies, suggesting that emerging districts may benefit from early investments in cultural amenities and aesthetic improvements, while established areas require ongoing attention to fundamental environmental quality and safety infrastructure.

The oscillating pattern in convenience perception highlights the importance of timing in public service provision. The temporary negative influence observed in 2021 suggests that rapid facility provision can create short-term disruptions to established street life patterns, but these effects can resolve over time as communities adapt to improved service provision. This finding supports phased approaches to service infrastructure development that allow communities to adjust to changes while maintaining existing social patterns.

4.4. Differentiated Policy Framework Based on Perceptual Thresholds

Based on our empirical findings, we propose a systematic framework for differentiated urban interventions that prioritizes specific perception dimensions according to spatial context and temporal dynamics. Our analysis reveals that different urban contexts require tailored approaches that acknowledge both the spatial heterogeneity of perception–vitality relationships and the nonlinear threshold effects we identified.

For central commercial districts, which exhibit high vitality but mixed perception performance, interventions should prioritize comfort and safety perception optimization. These areas typically show lower baseline performance in comfort due to higher density and building interface enclosure, suggesting that green infrastructure integration achieving 15–25% green view coverage and comprehensive pedestrian safety enhancements would generate the greatest vitality benefits. The threshold analysis indicates that comfort perception indices above 0.6 and safety perception indices above 0.7 are necessary to achieve meaningful improvements in these contexts.

Residential areas, characterized by moderate vitality and more balanced perception profiles, require a different approach focusing on convenience and pleasure perception enhancement while preserving existing comfort advantages. The inverse U-shaped relationship identified for convenience perception suggests careful calibration of mixed-use facility integration, targeting convenience indices of 0.8–1.2 to avoid oversaturation effects. Simultaneously, cultural amenity provision and aesthetic improvements can enhance pleasure perception, with target indices above 0.5 supporting sustained vitality growth.

Industrial and peripheral areas, which demonstrate consistently low vitality across all perception dimensions, necessitate comprehensive multi-dimensional improvements with safety and comfort as foundational requirements. Sequential intervention strategies should establish basic safety and comfort infrastructure with target indices above 0.4 before introducing convenience facilities and pleasure amenities. This approach acknowledges the threshold effects we observed, where improvements only generate significant vitality benefits after reaching moderate intensity levels.

The temporal evolution of perception priorities suggests that intervention strategies should adapt to urban development phases. Early development areas benefit from initial investments in pleasure perception to establish cultural identity and distinctive character, followed by basic convenience provision while avoiding early oversaturation. As areas mature, the focus should shift toward safety and comfort infrastructure that supports long-term livability, with convenience optimization occurring only after fundamental environmental quality is established.

Resource allocation strategies should leverage the threshold effects we identified by concentrating investments in areas closest to threshold achievement rather than distributing improvements thinly across large areas. This approach maximizes the likelihood of generating immediate vitality improvements that can demonstrate intervention effectiveness and justify expanded investment. The nonlinear relationships we discovered suggest that comprehensive intervention packages addressing multiple perception dimensions simultaneously are more effective than single-focus improvements.

4.5. Critical Reflections and Theoretical Contributions

Our findings necessitate critical examination of fundamental assumptions in urban planning theory and practice. The dramatic temporal reversals observed in perception–vitality relationships, particularly for safety perception, shift from negative to positive influence. This finding suggests that the relationship between built environments and human behavior is far more dynamic and contextually dependent than previously acknowledged, requiring fundamental reconsideration of how we conceptualize environmental effects in urban planning.

The identification of inverse U-shaped relationships, particularly for convenience perception, raises important questions about the assumptions of linear progress that dominate urban development discourse. The finding that moderate facility provision generates optimal vitality benefits while excessive provision diminishes street life contradicts conventional wisdom about maximizing amenity provision. This suggests that optimal urban environments may require carefully calibrated restraint rather than maximization of development intensity.

The spatial homogenization observed over time presents a theoretical paradox. While convergence toward standardized environmental effects might indicate successful urban integration, it also raises concerns about the loss of neighborhood distinctiveness and place-based identity that are fundamental to vibrant urban life. This tension highlights the need for planning approaches that can balance standardization benefits with place-specific character preservation.

Our methodological integration of MGWR and SHAP approaches reveals limitations in traditional urban analysis that relies on either purely spatial or purely statistical approaches. The complementary insights generated by combining spatial heterogeneity analysis with nonlinear machine learning demonstrate the inadequacy of single-method approaches for understanding complex urban phenomena. This methodological contribution challenges disciplinary boundaries between quantitative geography and urban planning, suggesting the need for more integrated analytical frameworks.

5. Conclusions

This longitudinal analysis of perception–vitality relationships in Hohhot demonstrates that the relationships between built environment perceptions and street vitality are fundamentally dynamic, spatially heterogeneous, and nonlinear. These findings challenge static approaches to urban planning and highlight the need for adaptive, context-sensitive design strategies.

Our research makes three primary contributions. First, we demonstrate that perception–vitality relationships undergo significant temporal evolution, with dramatic coefficient reversals and changing priority hierarchies over time. This challenges assumptions about environmental determinism and supports adaptive planning approaches. Second, our identification of nonlinear threshold effects and inverse U-shaped relationships provides crucial guidance for resource allocation, particularly the finding that moderate facility provision often generates optimal outcomes. Third, our methodological integration of MGWR and SHAP approaches provides a replicable framework for understanding complex urban phenomena while maintaining practical interpretability.

The implications for planning practice center on three key shifts: implementing spatially differentiated strategies that acknowledge local variations, adopting adaptive approaches with regular reassessment cycles, and developing integrated interventions that address multiple perception dimensions simultaneously. Our findings are particularly relevant for rapidly urbanizing cities, where early intervention windows may be critical for establishing positive perception–vitality relationships.

This research has limitations, including its focus on a single Chinese city and its cross-sectional temporal analysis. Future research should examine multiple cities with varying characteristics, integrate subjective perception surveys with objective measurements, explore the mechanisms underlying temporal evolution, and evaluate intervention effectiveness through experimental designs.

Ultimately, our findings from Hohhot suggest that successful urban planning requires embracing dynamic, context-sensitive approaches that acknowledge the complexity of human–environment relationships while remaining responsive to evolving community needs. The future of urban planning lies in adaptive, place-based approaches that can evolve with their communities and contexts.