1 Introduction
Historical districts are not only vital components of urban cultural heritage but also key sources of urban vitality (
Li et al., 2021). However, under the pressures of modernization and urbanization, these areas face dual challenges: they must fulfill the core mission of cultural heritage preservation while simultaneously responding to the economic demands of commercial development and industrial transformation. This contradiction has led to a widespread decline in street vitality (
Zhang et al., 2022;
Zhou et al., 2017). The decline in street vitality is especially pronounced in the old industrial bases of Northeast China, such as Harbin (
Xie et al., 2016). During the city’s development and transformation, historical districts have suffered from widespread vacancy and population outflow due to industrial restructuring. Traditional commercial areas like Daowai Historic District are experiencing both commercial homogenization and the erosion of historical character, posing a dual crisis. Revitalizing these memory-laden urban spaces has thus become an urgent challenge.
There is a growing academic consensus that street vitality is shaped by the dynamic interplay between human activity and the built environment (
Chen et al., 2022;
Jiang et al., 2022;
Tu et al., 2020). Accordingly, revitalizing historic districts requires a clear understanding of how the built environment and human behavior interact, calling for a systematic analysis of spatial factors and underlying mechanisms. Previous studies have proposed frameworks such as the “3D” and “5D” models (
Cervero and Kockelman, 1997;
Ewing and Cervero, 2010), emphasizing the role of density, diversity, design, and accessibility in shaping urban vitality (
Peng et al., 2023;
Xiao et al., 2021;
Yang et al., 2021). These frameworks have been widely applied to analyze street vitality in general urban settings, producing extensive empirical evidence (
Doan et al., 2025;
Li et al., 2021;
Xiao et al., 2021;
Yang et al., 2021).
However, the applicability of these findings to historic districts remains limited. Historic districts are not merely miniature versions of modern cities, they differ substantially in spatial morphology, functional constraints, and cultural perceptions. Spatially, historic districts often retain traditional street patterns characterized by narrow street widths, irregular networks, and compact building layouts with limited capacity for spatial transformation (
Bian et al., 2024). While these configurations may be considered inefficient in modern urbanism, they embody collective memory and cultural heritage (
Lyu et al., 2023). Functionally, the preservation policies imposed on historic structures constrain large-scale redevelopment, limiting the flexibility of functional adaptation (
Zhu et al., 2022). In some cases, excessive commercialization can lead to “touristification,” thereby undermining local identity and diminishing sreet vitality (
Wang et al., 2019). From a cultural perspective, the cultural and historical significance of such districts is not only a constraint on physical transformation but also an active force shaping how built environment elements are perceived and used (
Mohammad et al., 2013;
Shamsuddin and Ujang, 2008). For example, narrow or dim alleys may be perceived as “unsafe” in modern urban areas (
Jiang et al., 2017). But with appropriate lighting and design, it can be interpreted as “mysterious” and appealing within a historical context (
Wells and Baldwin, 2012). Such perceptual differences significantly shape spatial preferences and behavioral responses, ultimately influencing how street vitality is produced. Therefore, understanding street vitality in historic districts requires moving beyond general spatial logics and situating the analysis within the historical context.
Given these differences, existing research results on street vitality based on ordinary urban backgrounds face challenges in applicability in historical contexts. Specifically, current related research still has obvious deficiencies in the following aspects:
First, in recent years, machine learning methods such as XGBoost and SHAP have been increasingly used in urban vitality studies for their ability to capture nonlinear relationships and enhance model interpretability (
Peng et al., 2023;
Xiao et al., 2021;
Yang et al., 2021). Compared to traditional linear models like ordinary least squares (OLS) and geographically weighted regression (GWR), these approaches offer better insights into the complex mechanisms underlying vitality (
Xiao et al., 2021).
However, most existing applications remain focused on general urban areas, with limited attention paid to their applicability in historic districts. Given the distinctive spatial configurations, cultural attributes, and functional constraints of historic environments, the nonlinear characteristics and threshold effects of built environment variables may differ substantially from those in modern cities. For instance, while commercial clustering often enhances vitality in modern urban context (
Yang et al., 2021), excessive commercialization in historic districts can undermine cultural identity and, in turn, become a major factor in vitality decline (
Wang et al., 2019). Moreover, the heightened sensitivity of historic districts to environmental changes makes it difficult for traditional linear models to capture their dynamic patterns. Therefore, it is necessary to apply interpretable machine learning within the specific context of historic districts to investigate the nonlinear mechanisms of street vitality, in order to inform targeted conservation and renewal strategies.
Second, numerous studies have demonstrated that subjective perceptions of the built environment significantly influence individual spatial behaviors (
Lin and Moudon, 2010), which in turn shape street vitality. Solely relying on objective indicators often fails to capture the experiential dimensions of urban space (
Qiu et al., 2023). It is therefore essential to integrate physical attributes with subjective perceptions to better explain how vitality emerges. Especially in the context of historic districts, as carriers of collective memory and local identity, these spaces often evoke unique subjective perceptions, which in turn shape individuals’ spatial experiences and behavioral patterns (
Chan et al., 2024;
Jiang and Liu, 2024;
Mohammad et al., 2013). Investigating the role of subjective perception in shaping vitality thus contributes to a more holistic understanding of its underlying mechanisms and provides critical support for conservation and renewal efforts in historic districts.
However, traditional perception measurement methods, such as surveys and the Semantic Differential (SD) method (
Qiu et al., 2023;
Seiferling et al., 2017), are costly and have low spatial resolution (
Zhang et al., 2018), which has long hindered their integration into street vitality studies. Recent advances in computer vision and deep learning have provided new pathways for large-scale, fine-grained measurement of environmental perception (
Yao et al., 2019;
Zhu et al., 2024). The MIT Media Lab’s Place Pulse project used crowdsourced data to map six dimensions of human perception―“safety,” “beauty,” “depressing,” “lively,” “wealthy,” and “boring” (
Zhang et al., 2018). By integrating street view imagery (SVI) with this dataset and employing machine learning, large-scale perception prediction has become feasible (
Yao et al., 2019). This technological breakthrough offers a powerful tool for investigating the relationship between subjective environmental perception and street vitality, paving the way for more refined and comprehensive research in this field.
To address these research gaps, this study employs multi-source big data and machine learning techniques to quantitatively examine the nonlinear relationships and interaction effects of objective built environment characteristics and subjective perception on street vitality in historic districts. The study selects Daowai Historic District in Harbin as a case study. First, Baidu heatmap data is used to measure street vitality, while objective built environment factors and subjective perceptions are quantified based on the “5D” model and the MIT Media Lab’s Place Pulse project. Next, an XGBoost model and SHAP are applied to interpret and visualize the model results, analyzing the nonlinear impact of the built environment on street vitality and the interactive effects of key influencing factors. Finally, targeted built environment optimization strategies are proposed based on different street vitality typologies.
The contributions of this study are threefold: (1) It applies interpretable machine learning techniques within the specific context of historic districts to uncover the nonlinear mechanisms by which built environment factors influence street vitality, thereby addressing a research gap in existing studies that have largely focused on general urban areas; (2) It integrates both subjective and objective dimensions by leveraging deep learning and SVI data to finely measure perceived environmental qualities, revealing the critical role of environmental perception in shaping historical districts’ street vitality; (3) It proposes differentiated regeneration strategies based on the mechanisms of vitality formation, offering insights for urban planners and policymakers to better guide built environment interventions and promote the revitalization of historical districts.
2 Materials and methods
2.1 Study area
Harbin, a recognized historical city and a cradle of China’s industrialization. Currently, the city is undergoing a transition from a traditional heavy industry-based economy to a service-oriented and high-tech industrial structure (
Xie et al., 2016), raising the tension between heritage preservation and urban transition. This shift, along with changing demographics and social structures, has posed significant redevelopment and modernization challenges for many historic districts in the city. The Daowai Historic District (shown in Fig. 1), renowned for housing the nation’s largest cluster of Chinese Baroque architecture and its pivotal role in early industrial and commercial development (
Shao and Sun, 2023), epitomizes these dynamics. Despite its cultural significance, it has suffered from declining street vitality due to urban sprawl, functional obsolescence, and the erosion of historical character―a trajectory shared by heritage districts in cities such as Beijing’s hutong and Shanghai’s Shikumen neighborhoods. This case study focuses on Daowai not only for its architectural uniqueness but also for its representativeness of broader urban renewal dilemmas in China: balancing cultural identity preservation with vitality revitalization. By dissecting the Daowai case, this research aims to provide transferable insights for policymakers and planners navigating similar challenges in compact, heritage-rich Asian cities.
2.2 Study framework
Jan Gehl’s theory of interactional space suggests that street vitality arises from the fulfillment of individuals’ perceptual needs within street environments (
Gehl, 2011). Based on empirical social investigations, Jane Jacobs argued that the physical space of streets serves as the setting for interpersonal interactions and is the material foundation of street vitality (
Jane, 2016). Building on the social-spatial attributes of street vitality, Marcus interpreted it as a series of intangible activities shaped by the built environment (
Marcus, 2010). Collectively, these classical theories emphasize that subjective perception and objective spatial features are two key dimensions influencing street vitality, and together they form the foundational pathway through which the built environment impacts everyday urban life.
However, such theories often assume relatively stable and linear relationships between spatial form and vitality outcomes. In contrast, complex adaptive systems theory offers a dynamic and relational perspective by conceptualizing historic districts as evolving systems shaped by the continuous interplay of diverse, interacting components (
Shi et al., 2021). From this standpoint, street vitality in historic districts is not simply a function of discrete environmental variables, but rather an emergent property resulting from nonlinear, co-evolving interactions among multiple spatial, functional, and perceptual elements (
Batten, 2007;
Chen et al., 2024). These elements are interrelated through feedback loops, coupling mechanisms, and mutual adaptation (
Cutter et al., 2008;
Hernantes et al., 2019). This perspective enables us to see historic districts not merely as containers of activity but as complex socio-spatial systems in which vitality arises from dynamic configurations that defy simple linear explanations.
To integrate these insights, this study constructs a hybrid theoretical framework that combines classical theories of street vitality with the urban complex adaptive systems perspective. Specifically, we propose a “data-driven – mechanism analysis – decision response” research structure that unfolds in four stages: data collection, indicator extraction, machine learning-based nonlinear modeling, and SHAP-based interpretative analysis (Fig. 2). First, data on independent variables (built environment factors) and the dependent variable (street vitality) are collected and processed. Second, the XGBoost model is implemented in Python to capture nonlinear patterns in the data. Third, the SHAP method is applied to interpret and visualize the model results. Finally, targeted planning strategies are proposed based on different street vitality typologies. This integrated framework not only continues the tradition of humanistic urban theory by emphasizing residents’ perceptual experiences (
Duan et al., 2022), but also incorporates the principles of complex adaptive systems to explain the emergent, nonlinear dynamics of environmental elements. It thus offers a fresh analytical perspective for revitalizing street vitality in historical districts.
2.3 Data
(1) Geospatial Base Data: Geospatial base data primarily includes the road network, building outlines, building footprint areas, and building heights of Daowai Historic District. The road vector data is obtained from OpenStreetMap and undergoes georeferencing. After simplifying the roads in ArcGIS, we select sampling points at 30-m intervals, resulting in a total of 618 units. Building outlines, footprint areas, and building height data are sourced from Baidu Maps and transformed into the WGS-84 coordinate system using QGIS.
(2) Baidu Heatmap Data: this study adopts Baidu Heatmap data to measure street vitality. Baidu Heatmap data is collected by capturing users’ location information when accessing Baidu’s mobile applications. Compared to traditional methods, this data offers high temporal resolution, continuity, and broad spatial coverage, effectively reflecting spatiotemporal variations in population density (
Huang et al., 2023).
To ensure the objectivity, we crawled Baidu Heatmap data for the week from May 8 to May 14, 2023 (during the selected time period, there were no extreme weather conditions or major city events, eliminating uncontrollable external environmental influences to capture baseline activity patterns). Using 1-h intervals, we collected heatmap data from 8:00 a.m. to 9:00 p.m., representing street vitality. This resulted in a total of 98 samples (14 each day).The raw data in CSV format included geographic co-ordinates (WGS1984) and relative population density values. These were imported into ArcGIS Pro, projected to the local coordinate system, and processed through rasterization and reclassification. Finally, average density values were calculated to generate a spatial representation of street vitality across the study area.
(3) POI data: The POI data utilized in this study is sourced from the 2023 Amap Open Platform. Following the standards outlined in the “New Industry, New Format, and New Business Model Statistics Classification (2018)” and relevant literature, POIs are classified into eight categories: catering services, shopping services, accommodation services, insurance and financial services, lifestyle services, public facilities, transportation facility services, and sports and leisure services. This includes information such as the location, name, and type of each point. The data co-ordinates are then transformed into the WGS-84 coordinate system using QGIS.
(4) SVI data: In this study, street subjective perceptions and micro-scale built environment factors were extracted from SVIs obtained via the Baidu Maps. Considering both the spatial characteristics of historical districts and human visual perception, images were collected at 30-m intervals along the streets (
Wang and Xiu, 2023). A total of 618 sampling points were created, covering the entire Daowai historical district. At each sampling point, images were captured in four directions (0°, 90°, 180°, and 270°) with a tilt angle of 22.5° and a field of view of 90° (
Han et al., 2023;
He et al., 2023). This resulted in a total of 2472 image data points.
Additionally, semantic segmentation of SVIs was performed using a Fully Convolutional Network (FCN) and the ADK20K dataset (
Yao et al., 2019) to facilitate indicator calculations (Fig. 3). This dataset provides pixel-level annotations for 150 object categories relevant to urban environments, such as buildings, roads, vegetation, sky, vehicles, and pedestrians, making it highly suitable for built environment analysis. The proportion of each object at the sampling point is determined by calculating the sum of pixels for that color in images from all directions (
Jia and Zhang, 2021).
2.4 Methodology
2.4.1 External representation of street vitality based on Baidu Heatmap Data
This study extracts population aggregation intensity from Baidu Heatmap data to represent street vitality. Considering urban residents’ daily activity patterns (
Huang et al., 2023), the study classifies vitality data into weekdays and weekends. The average vitality from Monday to Friday represents weekday data, while the average from Saturday and Sunday represents weekend data. Street vitality intensity is quantified using the average heatmap value within specific time periods, calculated as follows:
In the equation, Vint represents the vitality intensity value; i denotes the ith moment; n is the number of moments involved in the calculation; and Vi signifies the heat value at the ith moment.
2.4.2 Objective features and subjective perception of built environment based on multi-source data
2.4.2.1 Objective features
Built environment objective features were selected based on the 5D built environment framework and a literature review (
Doan et al., 2025;
Han et al., 2024;
Liu et al., 2023;
Peng et al., 2023;
Wu et al., 2018). The spatial scope for calculating built environment variables was determined as a 55-m buffer based on prior studies (
Guo et al., 2021), which accurately reflects the functional distribution characteristics on both sides of the street. The 5D model proposed by Cervero describes the built environment through five dimensions: Density, Diversity, Design, Destination Accessibility, and Distance to Transit (
Cervero and Kockelman, 1997).
Density includes building density, historical heritage density, functional density, and the density of the three dominant commercial activities in the Daowai Historic District: Catering, Shopping, and Life services. Historical heritage density reflects the concentration of cultural heritage and is derived from the density of heritage and historic buildings within the street buffer.
Diversity is measured using the POI Shannon index, which quantifies functional mix and reflects the variety of land use types along the street (
Huang et al., 2020).
Design is evaluated through eight indicators extracted from SVI: Street Aspect Ratio, Color Index, Green view index (GVI), Enclosure, Sky View Index, Motorization Index, Walkability, and Signboards Density.
Destination Accessibility is assessed using closeness and betweenness calculated via Spatial Design Network Analysis (sDNA). Closeness centrality represents the ease of reaching other street networks within a given radius, indicating the built environment’s attractiveness to pedestrian flow (
Huang et al., 2023). Betweenness centrality measures a location’s potential as a travel corridor relative to other spaces (
Ma, 2020).
Distance to Transit is measured as the shortest distance from each sampling point to the nearest subway and bus station (
Chen et al., 2023).
The detailed calculation methods for each indicator are provided in Appendix A.
2.4.2.2 Subjective perceptions
Drawing on the Place Pulse 2.0 dataset from the MIT Media Lab (
Zhang et al., 2018), this study identifies six subjective perception indicators of the built environment: safety, beauty, depressing, lively, wealthy, and boring. Previous studies have demonstrated that these six indicators comprehensively represent human perception (
Dong et al., 2023;
Li et al., 2022;
Rui, 2023;
Zhang et al., 2018) and that variations in residents’ backgrounds do not introduce bias in the results (
Yao et al., 2019).
Following
Yao et al. (2019), we employed a machine learning-based scoring model to quantify subjective perceptions of the street environment. Specifically, we first selected 300 SVIs from the Daowai Historic District, ensuring that the sample distribution covered various street space types. Despite the relatively small sample size, prior research has confirmed that a dataset of this scale is sufficient to ensure data reliability (
Qiu et al., 2023).
Secondly, to more accurately assess the perceived street environment within the study area, we recruited 50 volunteers with local socio-cultural backgrounds to rate the SVIs based on six subjective perception indicators, using a human-machine adversarial scoring framework (
Yao et al., 2019). Ethics approval was obtained from the Ethics Committee of Harbin Institute of Technology. In addition, we have taken the participants’ permission and consent to participate in this study. The volunteer group was gender-balanced, aged between 23 and 60, and all participants had a solid understanding of the local context. Compared to expert panels used from (
Kexin et al., 2024;
Yao et al., 2019), which involved only 10 and 20 designers respectively, the size and the diversity of our expert panel is larger. Subjective perception ratings were categorized into five levels using the Likert scale: strongly disagree, disagree, neutral, agree, and strongly agree, corresponding to scores of 0–19, 20–39, 40–59, 60–79, and 80–100, respectively (
Cui et al., 2023). All volunteers received prior training on the scoring procedure and observed sample SVIs before evaluation. By computing the average perception scores of the sample images, we constructed the street perception dataset for the Daowai Historic District.
Finally, we extracted semantic segmentation features from the SVIs and trained XGBoost models using the manually scored perception dataset. To validate model performance, we applied 10-fold cross-validation and hyperparameter tuning. All models achieved R2 values above 0.70 and prediction accuracies exceeding 80%, indicating strong predictive performance and reliability (see Appendix D, Table D-1). The trained models were then used to estimate perception scores for all SVIs across the study area, generating a comprehensive dataset of spatially distributed subjective perceptions.
2.4.3 Modeling approach
This study adopts the XGBoost algorithm, an advanced implementation of gradient boosting decision trees (
Chen and Guestrin, 2016), to model the complex relationships between built environment variables and street vitality. XGBoost offers several key advantages: it excels at capturing nonlinear patterns and variable interactions, handles missing data effectively, and provides strong predictive performance even in high-dimensional settings (
Li and Managi, 2025;
Liu et al., 2022a,
b). These features are particularly suitable for urban vitality studies, especially in historical districts, where the interactions between spatial form, cultural context, and functional constraints are inherently nonlinear and context-dependent. Given these strengths, XGBoost was selected over other machine learning models as the core modeling technique. The model was implemented in Python 3.9 using the Scikit-learn library, with hyperparameter tuning conducted to optimize predictive accuracy.
2.4.4 Interpretation approach
To interpret the XGBoost model, we employed the SHAP package in Python 3.9. Originally proposed by Lundberg and Lee (
Lundberg and Lee, 2017), SHAP provides a robust and theoretically grounded framework for quantifying the contribution of individual variables to model predictions and detecting interaction effects among variables (
Iban, 2022;
Kim and Lee, 2023). Its ability to generate transparent and interpretable insights allows us to examine the marginal effects of each factor under varying contextual conditions and to understand how combinations of built environment attributes jointly influence vitality outcomes. These insights are particularly valuable for evidence-based urban planning and design.
3 Results
3.1 Spatiotemporal characteristics of street vitality
We observed that the fluctuation trends of street vitality on weekdays and weekends are generally similar, exhibiting a typical “rise-fluctuation-decline” pattern (Fig. 4(a)). However, compared to weekdays, street vitality on weekends demonstrates more pronounced fluctuations and higher peak levels. This observation corresponds with the results from
Yu et al. (2024), indicating that street spaces attract more diverse activities during weekends.
From a spatial distribution perspective, high-vitality streets are mainly concentrated in the southeastern part of the Daowai Historical District, where residential areas such as Taigu Xintiandi and NanXun Community are densely located. In contrast, low-vitality streets are primarily distributed in unrenovated historical areas, such as Tongfa First Street, North Second Street, and North Third Street (Fig. 4(b)). To further examine the spatial structure of street vitality, we conducted a global spatial autocorrelation analysis. The results show significant positive spatial autocorrelation for both weekday and weekend vitality, with Moran’s I values of 0.882 and 0.911, respectively. These values suggest that streets with similar vitality levels tend to cluster geographically rather than being randomly distributed (
Chen et al., 2023). Detailed methods and results are presented in Appendix D, Section 2.
Considering that weekday street vitality is predominantly influenced by constrained activities such as commuting (
Wang et al., 2023), this study selects weekend street vitality, which is primarily driven by spontaneous resident activities, as the dependent variable for constructing the analytical model.
3.2 Variable selection and model performance evaluation
Prior to applying machine learning algorithms, multi-collinearity tests were conducted to filter out ineffective features. Based on empirical research, a Variance Inflation Factor (VIF) below 7.5 indicates that the model does not suffer from multicollinearity issues (
Doan et al., 2025). The results showed that, except for Enclosure, all 23 remaining variables had VIF values below 7.5. Therefore, Enclosure was excluded. Appendix B provides the descriptive statistics and VIF test results for all variables.
To identify the most suitable model for nonlinear analysis, we constructed street vibrancy prediction models using Random Forest (RF), Gradient Boosting Decision Tree (GBDT), XGBoost (eXtreme Gradient Boosting), and LightGBM (Light Gradient Boosting Machine). The dataset was split into an 80% training set and a 20% test set, and model performance was evaluated by comparing training results. As shown in Table 1, XGBoost outperforms the other models across all evaluation metrics, achieving the lowest MAE (0.409) and RMSE (0.546), as well as the highest R2 (0.657). Compared to Random Forest, GBDT, and LightGBM, XGBoost demonstrates superior predictive accuracy and model fit, highlighting its strong capability in capturing nonlinear relationships and complex interactions among variables. Therefore, XGBoost was selected for further analysis.
To optimize model performance, K-fold cross-validation and grid search were used for hyperparameter tuning. After extensive testing, the optimal XGBoost configuration was achieved with: K = 5, colsample_bytree = 0.5, alpha = 1, n_estimators = 120, max_depth = 4, and learning_rate = 0.2, resulting in an R2 of 0.734, indicating optimal model performance.
Furthermore, to explore the contribution of built environment subjective perception to the prediction of street vitality in historic districts, we compared the performance of a model including subjective perception (Model 1) with a model excluding subjective perception (Model 2). As shown in Table 2, Model 1 achieved a lower RMSE (0.528) compared to Model 2 (0.597), with a reduction of 0.069. Additionally, Model 1’s R2 value (0.734) was 0.052 higher than that of Model 2 (0.682). These findings indicate that incorporating subjective perception significantly enhances the model’s explanatory power, allowing for a more accurate prediction of street vitality dynamics in historic districts.
3.3 Relative importance
Figure 5(a) ranks variable importance, while Fig. 5(b) groups them by category. Functional Density is the most influential factor, followed by Distance to Transit, Betweenness, and Functional Mix, highlighting the combined significance of service provision and spatial accessibility. Building Density and Historical Heritage Density also contribute notably, underscoring the value of spatial intensity and cultural assets in enhancing street vitality.
While subjective perceptions rank lower than objective features, their influence remains significant. Notably, Boring exerts the strongest negative effect, followed by Safety, Wealthy, and Lively. These perceptions shape users’ emotional and behavioral engagement with the street environment.
In the design dimension, GVI has the highest impact, supporting the role of visible greenery in historical district vitality, followed by Color Index and Signboard Density. Importantly, Distance to Transit ranks second overall, reinforcing the importance of transit-accessible design in balancing heritage conservation and urban connectivity.
Finally, Betweenness shows a much stronger effect than Closeness, contrasting with patterns in typical urban settings (
Yang et al., 2023). This may reflect the immersive experience encouraged by historic environments (
Svensson, 2021), where exploratory walking boosts vitality, in contrast to modern streets that prioritize travel efficiency (
Huang et al., 2023).
To verify robustness, we compared XGBoost and SHAP results, which show strong consistency. See Appendix D, section 3 for details.
3.4 Nonlinear relationships
This section employs partial dependence plots (PDPs) to analyze the influence of various factors on street vitality in historical districts. The PDPs reveal that the strength of each variable’s impact varies across different value ranges, indicating the presence of nonlinear relationships and threshold effects.
3.4.1 Subjective perception of the built environment
As shown in Fig. 6, Boring and Depressing are negatively associated with street vitality, and their inhibitory effects become more pronounced when the scores exceed 56 and 52, respectively. Safety shows an inverted U-shaped pattern, with a positive effect emerging at a score of 37, peaking at 42, and then gradually declining. A similar pattern is observed for Beauty, which begins to have a positive effect at 43, reaches its peak at 49, and then levels off. In contrast, Lively and Wealthy perceptions are positively correlated with street vitality, with local effects becoming positive when scores exceed 38 and 52, respectively.
3.4.2 Objective characteristics of the built environment
As illustrated in Fig. 7, in the Density dimension, both Functional Density and Building Density are positively associated with street vitality. The effect of Functional Density becomes positive beyond a value of 5, while Building Density turns positive after 0.4 but shows diminishing returns beyond 0.5. Historical Heritage Density follows a U-shaped pattern, with significant positive effects emerging only when density exceeds 20. Shopping, Catering, and Life Service Density display inverted U-shaped relationships, indicating vitality gains up to a certain threshold, followed by decline.
In the Diversity dimension, Functional Mix peaks at 1.15 before declining, suggesting an optimal range for land-use diversity. This nonlinear trend may clarify conflicting findings in previous linear-model studies (
Chen et al., 2022;
Tu et al., 2020). Distance to Transit shows the strongest positive effect around 130 m, but its influence weakens sharply beyond 250 m and becomes negligible after 300 m.
In the Destination Accessibility dimension, Closeness is positively associated with vitality, becoming effective when it reaches 260. Betweenness displays an inverted U-shaped relationship, peaking at 5 and declining after 9, which implies that moderate pedestrian network centrality enhances vitality, while excessive fragmentation may hinder it.
In the design dimension, GVI and Walkability positively correlate with street vitality, consistent with previous studies (
Liang et al., 2022;
Zhang et al., 2024). Color Index, Street Aspect Ratio, Sky View Index, Motorization Index, and Signboards Density all exhibit an initial increase followed by a decline, suggesting optimal thresholds.
3.5 Interaction effects
Figure 8 illustrates interaction effects among the six most influential variable pairs, focusing on how combinations of built environment features shape street vitality. The strongest synergy is observed between Functional Density and Distance to Transit. When transit distance is under 250 m and functional density exceeds 8, street vitality increases markedly. Interactions between Density and Design dimensions also emerge. High GVI tends to weaken the positive impact of both Functional Density and Building Density. Building Density and Betweenness reinforce each other when both exceed critical thresholds (0.3 and 12.5, respectively). A similar pattern is found between Street Aspect Ratio and Betweenness, where wider streets (ratio >1.2) can amplify the effects of high connectivity. Lastly, Distance to Transit interacts positively with Safety: when stations are within 250 m and perceived safety scores exceed 45, vitality improves significantly.
3.6 Formation mechanisms of street vitality across different types
To reveal differentiated mechanisms shaping street vitality, we performed hierarchical clustering based on the similarity of local effects from built environment variables. Three distinct street types emerged (Fig. 9): high-vitality―transit-accessible streets (27%), moderate-vitality―density-constrained streets (31%), and low-vitality―design-compensated streets (42%).
High-vitality―transit-accessible streets are mainly located along the periphery and the southeastern residential zones of the district. Convenient public transit plays a crucial role in enhancing vitality in these areas. Additionally, high Building Density and Functional Density further contribute to the clustering of street vitality.
Moderate-vitality―density-constrained streets are predominantly found in the middle sections of internal streets, where historical buildings are concentrated. Despite well-preserved architectural heritage, the lack of functional facilities limits their ability to attract vibrant street activity. In addition, low street betweenness further suppresses vitality by restricting pedestrian movement.
Low-vitality―design-compensated streets appear mostly at internal intersections with limited connectivity and access. While structurally constrained, these streets show improvement potential through design interventions, such as increasing greenery (GVI), enhancing visual color contrast, and adjusting signage layout to support walkability and visual interest.
4 Discussion
4.1 The nonlinear relationship between street vitality and the built environment in historical districts
Our findings align with existing literature, demonstrating a nonlinear relationship between the built environment and street vitality. Building Density positively correlates with street vitality, with a notably stronger association when exceeding 0.4. However, the correlation plateaus beyond 0.5, which could be explained by the tendency of high density to reduce open space, thereby coinciding with lower vitality (
Xiao et al., 2021). Similarly, Historical Heritage Density density and service facility densities exhibit threshold effects, suggesting that excessive single-use clustering correlates with suppressed vitality.
Interestingly, our study find that historic districts exhibit unique threshold effects and nonlinear patterns compared with studies in urban contexts. For instance, we identify an inverted U-shaped relationship between Functional Mix and street vitality, similar to subway station areas (
Xiao et al., 2021). However, the threshold (1.15) in historical districts is significantly higher than that in subway station areas (0.8) (
Xiao et al., 2021). This discrepancy may stem from functional differences: subway stations primarily serve commuting functions, where user behavior is goal-oriented and time-sensitive (
Jiao et al., 2023), whereas historical districts function as cultural consumption spaces, fostering prolonged engagement and diverse experiences (
Wang et al., 2015). Thus, historical districts require a higher degree of Functional Mix to accommodate diverse needs (
Huang et al., 2023).
Furthermore, our findings show that three-dimensional built environment factors―such as Street Aspect Ratio, Sky View Index, and Signboard Density―exhibit inverted U-shaped relationships with street vitality, in contrast to the mostly monotonic trends observed in broader urban contexts (
Wang et al., 2023;
Yang et al., 2023). This suggests that built environment factors in historical districts have explicit threshold effects, beyond which their marginal benefits reverse. Compared to general urban areas, street vitality in historical districts is more sensitive to spatial form, highlighting the need for “moderate and orderly” design.
These findings support our hypothesis that the distinct cultural heritage, industrial structure, spatial layout, and architectural style of historical districts result in unique street vitality mechanisms. These results suggest that urban-scale strategies may not be directly applicable urban-scale strategies are not directly applicable to historical districts, necessitating refined, context-specific revitalization strategies.
4.2 Interaction effects of street vitality factors
Our research finds that certain combinations of built environment features are associated with higher levels of street vitality, particularly within specific threshold ranges. These associations suggest potential directions for targeted design interventions to improve street vitality. For instance, when Distance to transit is within 250 m and Functional Density exceeds 8, we observe amplified vitality levels (Fig. 10). This aligns with the “flow-function” reinforcement mechanism: public transport ensures a stable influx of people, while dense functional facilities provide activity spaces, prolonging stay duration (
Tang and Ta, 2022). Therefore, historical district revitalization should adopt a “transport accessibility þ functional density” dual-strategy approach to leverage their synergistic association with vitality.
Similarly, the co-presence of high building density (>0.3) and high betweenness (>12.5) correlates with significantly elevated vitality. High density accommodates commercial and residential functions, while Betweenness optimizes pedestrian connectivity. A comparable pattern emerges with street aspect ratio (>1.2) and betweenness (>7). These findings underscore the importance of coordinated “morphological control” and “network optimization” in historical district renewal, aligning with urban system synergy theories (
Chen et al., 2024).
Moreover, the results reveal that high functional/building density combined with elevated GVI correlates with reduced vitality, contrasting with conventional greenery benefits (
Yang et al., 2023). This could be due to the displacement of essential commercial interfaces and activity spaces in high-density zones (
Bibri et al., 2020). Therefore, strategic greening approaches (e.g., vertical integration) warrant consideration where density constraints exist.
Furthermore, Distance to transit moderates safety’s association with vitality: Near transit (<250 m), higher safety perception (>45) correlates with significantly greater vitality gains than in distant areas. This pattern indicates that safety perceptions may play a heightened role in transit-adjacent zones (
Ingvardson and Nielsen, 2022).
4.3 The impact of subjective perception
Although subjective perception can comprehensively reflect individuals’ cognition of the environment, it has been underutilized in previous studies (
Dong et al., 2023). This study integrates deep learning and SVI to reveal the relationship between subjective environmental perceptions and street vitality in historical districts, with a particular focus on their nonlinear patterns.
First, our findings demonstrate that incorporating subjective perceptions enhances the predictive accuracy of our model, reinforcing the necessity of considering both aspects (
Zhu et al., 2024;
Dong et al., 2023). Among subjective perception factors, boring shows the strongest negative correlation with vitality, aligning with studies on environmental perception and walkability (
Zhu et al., 2024), followed by significant associations for safety, wealthy, lively, beauty, and depression.
A deeper analysis of the nonlinear relationships between subjective perceptions and street vitality reveals that boring and depression are negatively correlated with vitality. Excessive negative spatial perceptions trigger physiological stress responses, reducing users’ willingness to linger and engage in activities (
Ulrich et al., 1991). Interestingly, in contrast to Jacobs’ emphasis on visual diversity and “eyes on the street” as drivers of vitality (
Jane, 2016), we find that excessive aesthetic and safety scores correlating with reduced vitality. Specifically, safety perception follows an inverted U-shaped pattern, peaking at a score of 42. This suggests that the enhancement of safety perception must remain within a reasonable range; once a general sense of safety is achieved, further improvements yield diminishing or even negligible effects. Similarly, beauty perception follows a comparable trend, reaching a peak before stabilizing. Excessive aesthetic appeal is not necessarily associated with increased street vitality, as
Zhang et al. (2018) found that aesthetic perception is primarily related to natural elements rather than artificial structures. Thus, overly aesthetically dominant streets often lack social interaction spaces, reducing activity clustering probability.
Lively is positively associated with actual street vitality, confirming the alignment between subjective and objective vitality levels. Similarly, wealthy also exhibits a positive correlation with street vitality. When wealthy perception is high, users tend to develop positive beliefs about the economic status of the environment, enhancing spatial identity and activity participation. This confirms the close relationship between socioeconomic vitality and spatial vitality from a subjective perception perspective (
Liu et al., 2022a,
b).
4.4 Implications for planning practice
Our findings provide several key insights for the revitalization of historic districts and policy formulation.
Firstly, planners must recognize that street vitality depends on both objective built environment features and subjective perceptual experiences. Therefore, planning strategies should explicitly address perceptual enhancement. For example: Within a 250-m radius of transit stations, urban designers should improve perceived safety through pedestrian-oriented lighting, clear sightlines, and wayfinding signage to foster pedestrians’ sense of security, which in turn may support higher levels of street activity.
Secondly, planners must acknowledge the unique nonlinear relationships and sensitivity thresholds of built environment variables in historic districts. It differs substantially from that in other urban streets, particularly regarding three-dimensional spatial factors, which exhibit higher sensitivity to specific thresholds. Therefore, planners cannot directly apply strategies from general urban streets to historic districts but must determine optimal values for built environment features based on their unique characteristics to effectively enhance vitality. For example, maintaining a functional mix index around 1.15 ensures that POI diversity contributes optimally to street vitality. Likewise, ensuring betweenness centrality remains within a moderate range (5–9) can enhance pedestrian flow and engagement. Overly complex or fragmented networks, by contrast, may hinder walkability and reduce street-level vitality. To support application, Appendix C summarizes optimal thresholds across all key variables, offering practical guidance for heritage-sensitive spatial interventions.
Additionally, this study underscore the necessity for urban planners to move beyond single-variable considerations in historic district revitalization. Instead, planners should account for threshold effects among multiple variables to avoid negative interactions while maximizing synergistic benefits (
L. Z. Xiao et al., 2021). For instance, interventions should not only improve transit accessibility but also ensure sufficient functional density around transit nodes to activate pedestrian flows and promote longer street-level engagement. Likewise, greenery should be distributed with attention to building and functional density―using vertical greening, planter boxes or movable planters in high-density areas to maintain visibility and pedestrian permeability.
Finally, the mechanisms underlying street vitality vary across different street types, necessitating differentiated strategies for vitality enhancement. Based on the classification of street vitality mechanisms, moderate-vitality―density-constrained and low-vitality―design-compensated streets exhibit relatively low vitality levels and should be prioritized for planning interventions. As shown in Fig. 11(a), for moderate-vitality―density-constrained streets, functional density can be enhanced through the adaptive reuse of historic buildings, such as incorporating cultural exhibitions, intangible heritage workshops, and other light-touch functions. In addition, activating ground-floor spaces through measures like outdoor commercial seating can enhance the attractiveness of functional facilities. Furthermore, improving street Betweenness to 5―for example, by opening internal courtyards to create additional pedestrian connections―can increase both pedestrian flow and the functional utilization of historic buildings. As shown in Fig. 11(b), for low-vitality―design-compensated streets, vitality can be enhanced through interface optimization: introduce layered greenery to increase the green view index to 0.1, such as low shrubs for enclosure, mid-height planters to guide pedestrian flow, and vertical greening on blank walls; implement color schemes that align with the historic character to adjust color richness to 31; and regulate commercial signage by controlling its size and clustering, maintaining a signboard density of 0.04.
4.5 Limitations
Undoubtedly, this study has limitations. Firstly, this study takes the Daowai Historic District in Harbin as a case. Although the single-case study approach may somewhat limit the generalizability of the findings, the district serves as a representative example of typical challenges faced by historical districts in Northeast China. The mechanisms revealed through in-depth analysis of this representative case hold broad theoretical and practical relevance (
Li et al., 2021). Future research should conduct cross-regional, multi-case comparative studies to enhance the applicability of the findings. Secondly, this study focuses on baseline street vitality during a single week, without accounting for temporal variations arising from seasonal changes or irregular events. Although the selected period was intentionally controlled to reflect typical activity patterns, future work should incorporate extended temporal datasets across different time frames and conditions to capture the dynamic nature of street vitality more comprehensively. Moreover, consistent with previous research (
Li and Managi, 2025;
Xiao et al., 2021), the use of interpretable machine learning methods to identify nonlinear patterns and interaction effects of built environment factors should be understood as correlational rather than causal. The interpretations of interaction effects and threshold mechanisms in this study are intended to propose research hypotheses and inspire future theoretical and empirical work, rather than to establish definitive causal pathways.
5 Conclusion
Based on the comprehensive analysis and discussion presented above, several conclusions can be drawn from this study.
First, subjective perceptions significantly shape street vitality: boring and depression suppress vitality, while a lively or beauty atmosphere enhances it. However, overly high scores in safety or beauty may paradoxically reduce vitality. Integrating these perceptions with objective built environment features offers a more comprehensive understanding of street vitality patterns.
Second, the relationship between built environment characteristics and street vitality is nonlinear, with clear threshold effects that differ markedly from those observed in contemporary urban contexts. For instance, vitality peaks when the Functional Mix Index approaches 1.15 and betweenness centrality ranges between 5 and 9. Planners should not simply apply strategies from general urban streets to historic districts but should tailor built environment configurations to their unique characteristics.
Moreover, interaction effects are observed among various variables, such as between functional density and distance to transit. Understanding these synergies enables designers and planners to optimize element combinations, thereby enhancing vitality through fine-grained spatial interventions.
Lastly, vitality mechanisms vary by street type: high-vitality―transit-accessible streets benefit from transit access and functional clustering, while low-vitality―design-compensated streets streets require targeted design interventions. We advocate for differentiated and context-specific strategies to enhance vitality based on street typology.
Overall, this study addresses the pressing challenge of declining street vitality in historical districts. By employing interpretable machine learning techniques, it systematically integrates subjective perceptions with objective built environment features to uncover the multidimensional mechanisms shaping street vitality. The findings offer theoretical and practical insights for revitalizing historical districts, with broader relevance for similar areas across Asia amid urban transformation.
2095-2635/2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.
PDF
(9187KB)
