Nonlinear effects of the built environment on urban vitality in Jinan based on multi-source data and explainable AI

Abstract

Understanding the nonlinear relationship between the built environment and urban vitality is a key challenge in urban planning. This study proposes an integrated analytical framework: (1) Combining multi-source data with explainable AI (XGBoost-SHAP) to decode nonlinear relationships and threshold effects; (2) Using the CRITIC weighting method for the first time to construct a composite urban vitality index; (3) Establishing a two-dimensional built environment framework using Baidu heat maps, nighttime light data, POIs, and street-view imagery to integrate “objective form” and “subjective perception”. The findings reveal: (1) Urban vitality shows a distinct “east-strong, west-weak” clustering pattern; (2) Objective factors dominate, especially POI blending degree, density and spatial integration, while perceptual factors exhibit limited influence; (3) Key nonlinear thresholds are identified: POI blending best promotes vitality above 1.5; density shows diminishing returns beyond 4,000 units/km²; and openness follows an inverted U-shaped curve. This study promotes the methodology for measuring comprehensive urban vitality, enriches the dimensions of the traditional framework of built environment indicators, and provides a new analytical framework for planning in historically layered cities.

Introduction

The concept of urban vitality first appeared in Jane Jacobs’ book The Death and Life of Great American Cities, published in 1961. Jane Jacobs argued that vibrant urban life fundamentally depends on neighborhoods with a diversity of activities, mixed uses, small streets, and high pedestrian flow¹. According to Ewing and Cervero, quality development and people-centered planning are critically important in the context of global urbanization. In this context, enhancing urban vitality has become a cornerstone objective of contemporary urban governance. Because of this, residents’ quality of life and spatial satisfaction is enhanced. According to Landry²and Wu³, urban competitiveness is reinforced through attracting talent, capital, and innovation. Advanced smart city technologies and multi-source spatial data have emerged. These include street view imagery, nighttime light remote sensing, and population heat maps. They have facilitated a deeper examination of the links between the built environment and urban vitality^4–6.

The interaction between the built environment and urban vitality has been explored across four orientations: objective built environment indicators, micro-perceptual indicators, ways of measuring vitality, and nonlinear modeling methods. For the objective indicators, many studies used the 5D framework proposed by Ewing⁷, which includes density, diversity, design, and destination accessibility. Indeed, studies underline the primacy of the blending degree and density of POIs in determining key drivers of vitality. As an example, in Wuhan, Li⁸applied XGBoost to analyze vitality. The results showed that POI mixes and road integration explained over 40% of network vitality. These two factors represent functional diversity and spatial accessibility. In the same vein, Jin⁹showed that POI density and the floor area ratio substantially increased the economic and social vitality of Beijing by enhancing the frequency of human activities. However, these studies focus primarily on spatial “physical attributes.” They rarely consider human subjective perceptions of space. This creates a gap between “what space offers” and “how people feel in it.” Recent advancements in street view imagery and computer vision have enabled the extraction of micro-level perception indicators¹⁰ Examples include green landscape coverage, spatial openness, and walkability. Unlike purely physical measures, these indicators capture human-scale spatial experiences. For instance, Xia¹¹ found that the green view index positively associated with street vitality across fifteen Chinese cities, as visible greenery improved pedestrian comfort. Moreover, according to Qin¹², a threshold effect was present in openness: moderate enclosure like in historic alleys increased social interaction, while large-open types of space like in large plazas instead suppressed it. However, perceptual indicators remain inadequately represented in most studies, often limited to a secondary role or a few isolated variables. Consequently, this limitation hinders a systematic comparison between objective and subjective influences, as well as the exploration of their synergistic effects. Measurement of vitality has progressed from field studies to multi-source data integration. Generally, social vitality is measured with Baidu Heatmap data¹³, economic vitality with nighttime light data¹⁴ or enterprise POI density¹⁵, and cultural vitality with cultural facility POIs¹⁶. However, this remains limited in that it calls for dimensionally singular indicators. For example, some studies focus solely on heatmap data, neglecting the symbiotic relationship between economic and cultural activities¹⁷. Several composite indices have been constructed. However, many use relatively simple weighting methods—such as the equal-weight average approach. These methods fail to reflect the differentiated contributions of each vitality dimension. In early analytical models, such as Ordinary Least Squares and Geographically Weighted Regression, relationships were presumed to be linear¹⁸. However, these models did not account for nonlinear dynamics, such as threshold effects or diminishing returns. Researchers have adopted machine learning models (e.g., XGBoost) to address these complexities. For example, Liu¹⁹ identified a critical threshold effect of population density on Shanghai’s vitality. Recently, the interpretable part of machine learning was brought by introducing the SHAP framework, which allows information on the contribution of each variable to be quantified²⁰. Most studies have not examined how objective and subjective factors interact. This makes it hard to fully understand how the built environment affects vitality.

Three significant research lacunae still pertain, especially in Jinan, where the historical and modern contexts coexist. The first remains the systematic integration and comparison of objective and subjective indicators. In this respect, the bias of prioritizing physical infrastructure over perception leaves many questions unanswered: Can areas with high public facility density but poor walkability maintains their vitality? Secondly, there remains a relative underdevelopment in the measurement of vitality in its entirety. Most literature considers one dimension of vitality, thus ignoring possible synergy or trade-offs between social and economic and cultural vitality. Aggregation methods are commonly used, but they often disregard the relative importance of different dimensions. Furthermore, as mentioned earlier, most linear models (and even some machine learning applications) fail to capture variable interactions and threshold effects. For instance, they do not explain how POI blending degree and spatial integration jointly influence vitality. Jinan seamlessly integrates historic districts (such as Quancheng Road-Daming Lake), modern central business districts, and emerging suburban areas. This blend makes it an ideal case study for examining the relationship between urban spatial characteristics and urban vitality across diverse built environments. For example, historical areas have low building density but high perceptual comfort, while CBDs have high POI density but low openness. Past research on Jinan has mostly studied single areas or separate indicators. Some only looked at basic relationships between urban elements. These studies did not explore how objective and subjective factors interact. To address the identified gaps, four specific research questions are proposed: (1)What are the spatial patterns of social, economic, cultural and comprehensive urban vitality in Jinan’s central urban area? (2) How do objective and subjective built environment factors nonlinearly influence urban vitality and what are the critical thresholds? (3) What are the relative contributions and synergistic effects of objective functional indicators and subjective perceptual indicators on urban vitality? (4) What context-sensitive spatial planning strategies can be proposed to enhance urban vitality in Jinan, considering its mixed historical and modern urban fabrics?

Materials and methods

Research area

This research thus confines itself to the central urban area of Jinan, as shown in Fig. 1, with an area of 1,533.66 km² and borders set according to the 2021–2035 territorial plan for the city. It covers parts of the Lixia, Shizhong, Huaiyin, Tianqiao, and most of the Licheng districts. Here live 3.92 million residents, which is 42% of the total population of Jinan. This area is predominantly demographic, economic, and administrative and well summarizes the deep historical legacy and modern development of the city. Using this in-depth analysis of built environment-vitality relationships is justified. Using road-defined blocks as spatial units, the study creates 2,276 units for fine-grained analysis, ensuring objectivity through natural boundaries.

Fig. 1.

Location of the central urban area of Jinan, China.

Data sources and preprocessing

Single vitality data

Social vitality (SV) was derived from Baidu Heatmap data. A week with no special weather or holiday influence was selected, covering August 5–9, 2024. The grid size of the data is 170 m × 170 m, and the time range from 07:00 to 22:00 represents the main commuting and leisure times. The information is counted every hour, and outliers above three times the average heat value are eliminated. The final index was calculated as the average heat value for 105 valid hours.

Economic vitality (EV) was measured by combining nighttime light with the density of enterprise POIs. The daytime density of key economic actors clustered in enterprises reflected by enterprise POIs^21,22 and captured daytime economic activities, such as office work and retail, helps to supplement nighttime light data, which is occasionally subject to local interference^23,24. The enterprise POIs were obtained from Gaode Map 2024 and classified according to the “Classification of National Economic Industries,” with eight categories concerning enterprise activity retained (e.g., wholesale/retail, finance). A total of 26,403 valid POIs were identified in this work to calculate their density. Nighttime light data sourced from the National Earth System Science Data Center (http://www.geodata.cn) had a 1 km resolution, which was then resampled using the nearest-neighbor method. An overall EV index was created using the entropy weight method to reduce single-source bias and better reflect the spatial economic patterns.

It was operationalized by the density of POIs related to cultural facilities, based on data sourced from Gaode Map 2024. This covers museums, libraries, galleries, theaters, convention centers, and so on. After screening and duplication, a total of 15,890 POIs were retained, and their density in each unit of analysis was computed to serve as an indicator of CV.

Built environmental data

Street view images from the Baidu Street View Map were used to capture subjective perceptual data. Sampling points were set every 50 m along the road network using the Baidu Map API, with four horizontal angles—0°, 90°, 180°, and 270°—taken at each sampling point. After removing invalid points, to ensure the representativeness of street view sampling, 7400 valid sampling points (corresponding to 29,600 images) covered all types of streets within the study area, including primary roads secondary roads, tertiary roads, trunk, and residential roads. The sampling density (one sampling point every 50 m) meets the fine-grained analysis requirements of urban blocks. The four horizontal angles of each sampling point ensure comprehensive coverage of street spatial features. These were processed using the DeepLab-V3Plus model²⁵, which is a 125-layer CNN trained with the Cityscapes dataset, for 19 classes of street view elements, such as people, vehicles, roads, and plants, to segment and calculate pixel proportions for each class. The network structure is shown in Fig. 2.

Fig. 2.

Network structure of the DeepLab-V3 + semantic segmentation model.

Objective built environment data were compiled from several sources: POI density, POI blending degree, and bus stop density from Gaode Map; building density and road network data from OpenStreetMap; population density from WorldPop, cross-validated with the Seventh National Population Census; and subway station locations from Jinan Metro’s official website. Detailed variable calculations are summarized in Table 1.

Table 1.

Calculation formula for built environmental indicators.

Dimension	Category	Variable	Illustrate
Objective	Density	Building density	Sum of built-up area within the unit/Unit area
		Population density	Sum of population within the unit/Unit area (persons/km2)
	Diversity	POI density	Total count of POIs within the unit/Unit area (units/ km2)
		POI blending degree	Where denotes the information entropy, and represents the proportion of the i-th category of POI in the unit.
	Transportation accessibility	Density of bus stops	Sum of bus stops within the unit/Unit area. (units/ km2)
		The distance to the nearest subway station	The distance from the centroid of a spatial unit to the nearest subway station. (m)
	Destination accessibility	Integration	represents the integration degree value of link line x, and represents the collection of broken lines within the radius network starting from link line ; represents the weight of broken line , represents the proportion of any broken line in the radius, represents the distance from the starting polyline to the ending polyline
		Choice	represents the intermediate index of ; represents the shortest path between and through the polyline ; represents the total weight value with each as the radius. represents the collection of polylines within the radius network starting from the link line . represents the weight of broken line , represents the proportion of any broken line in the radius
Subjective	Perception	Green view index(GVI)	, Where represents the pixel ratio of the green plant elements in the j-th image, and n represents the number of sampling points within the spatial unit
		Openness(O)	Where represents the pixel ratio of the sky in the j-th image.
		Walkability(W)	represents the pixel ratio of the sidewalk in the j-th image, represents the pixel ratio of the fence in the j-th image, and represents the road pixel ratio in the j-th image.
		Imaging property(I)	represents the pixel ratio of the building in the j-th image, represents the pixel ratio of the Signs in the j-th image.

Spatial-temporal alignment and fusion of multi-source data

To address the temporal inconsistency of multi-source data, this study adopts a strategy of “unified spatial standards + target fusion”. All data are aggregated and associated using research units (2,276 blocks) generated based on the road network and administrative boundaries in 2023 as a unified spatial carrier, ensuring the spatial comparability of all indicators. First, for the fusion alignment of single-source data:

The street view images are from Baidu Street View, with a sampling time of March 2023, covering multiple seasonal features and showing no significant seasonal differences. Street view data remains highly stable in the short to medium term (2–3 years) in the absence of large-scale renovations. Therefore, it is reasonable to associate it with the recent vitality data. POI data was obtained through Gaode Maps in 2024, and invalid or duplicate POIs were deleted through spatial overlap analysis. The sampling period of Baidu’s heat map data was from 07:00 to 22:00 on August 5th to 9th, 2024. The average value was used to represent the stable state of population activities rather than instantaneous fluctuations. Nighttime lighting data: the annual synthetic data of the National Earth System Science Data Center (2023), with an annual scale, can effectively smooth out short-term fluctuations (such as holidays) and more stably reflect the intensity level of regional economic activities. In this study, the nearest neighbor method was used to re-sample to 170 m×170 m (matching the heat map resolution). Integrate the poi of enterprises in 2024 with that of 2024 through the entropy weight method. The population data source used was the 2020 WorldPop dataset, which was cross-verified with the Seventh National Census (2020). The 2020 data from WorldPop was the latest high-resolution population dataset with global consistency that was publicly available at the time of the design of this study. This provides a reliable population distribution benchmark for the research.

Multi-source data fusion method: Economic vitality (EV) integration: By integrating nighttime light data (nighttime economic activities) and enterprise POI density (daytime economic activities) through the entropy weight method. The integration of comprehensive urban vitality (COV): Social (SV), economic (EV), and cultural (CV) vitality is integrated through the CRITIC weighting method, which takes into account the variability and conflict of indicators (avoiding subjective bias in equal-weighted averaging). Integration of building environment indicators: Objective/subjective indicators are standardized to [0,1] through minimum-maximum normalization, and then input into the XGBoost model for unified analysis to eliminate dimensional interference. In conclusion, through the comprehensive processing of data sources, the problem of weakened perception indicators caused by data asynchrony has been minimized to the greatest extent, ensuring the reliability of the data in this study.

Methods

Following a structured technical roadmap (Fig. 3), this study first maps and integrates three vitality dimensions into a Composite Urban Vitality (COV) index for Jinan. We then analyze their spatial patterns and define the built environment through twelve key indicators. The core analysis applies the XGBoost-SHAP framework to unravel the relationship between COV and the built environment, directly informing the subsequent proposal of spatial planning strategies.

Fig. 3.

Technical roadmap of the research methodology.

Using the CRITIC weighting method to construct the comprehensive vitality

The CRITIC method (Diakoulaki 1995) objectively assigns weights to indicators by considering both their contrast intensity (variability) and conflict (correlation). This approach outperforms entropy and average weighting by leveraging inherent data properties, avoiding the bias that higher values imply greater importance. It is particularly suited for urban vitality systems, where indicators exhibit significant correlations and scale differences. The computational steps are as follows:

Given evaluation objects and evaluation indicators, a data matrix can be constructed. Let the standardized elements of the matrix be denoted as . First, standardize the elements.

1

Second, calculate the contrast intensity of the -th indicator, where represents the standard deviation, is the sample size, and is the mean value of indicator .

2

3

Next, compute the conflict index of the evaluation indicators, where is the sample size. The conflict index reflects the degree of correlation between different indicators; a smaller value indicates stronger positive correlation. Let denote the magnitude of contradiction between indicator and the remaining indicators.

4

5

Here,represents the Pearson correlation coefficient between indicator and indicator . Then, calculate the information load capacity .

6

Finally, compute the weight and the comprehensive vitality score.

7

8

Spatial autocorrelation

We used spatial autocorrelation analysis to measure the distribution patterns of spatial elements in the study area and their statistical correlations, including global and local autocorrelation. Global spatial autocorrelation reflects the overall distribution trends of spatial objects within the study area. Local spatial autocorrelation is used to identify the local characteristics of spatial object distribution within the study area. We selected Moran’s I and Getis-Ord G* as the autocorrelation statistics for global and local analysis respectively. For the spatial weight matrix, we chose the “Queen’s adjacency” pattern:

9

10

Among them,, the element is the deviation of the attribute value of element from its average value , is the spatial weight between element and , is equal to the total number of elements, and is the aggregation of all spatial weights, .

XGBoost + SHAP model

Extreme Gradient Boosting (XGBoost) is a powerful machine learning algorithm for classification and regression^26,27. Building on Gradient Boosting Decision Trees, it delivers higher accuracy than methods like SVM and Random Forests. XGBoost improves prediction by integrating multiple trees, handles multicollinearity without distributional assumptions, and resists outlier influence. Its objective function combines a loss function with a regularization term, expressed as:

11

In the formula, , . Here, is the penalty coefficient for the number of leaf nodes, which is used to constrain model complexity and reduce overfitting; represents the number of leaf nodes; λ is the penalty coefficient for leaf node scores, which controls the regularization strength of leaf node scores in the model; and denotes the score of the -th leaf node. After optimization through grid search and Bayesian optimization, the optimal model parameters were determined (Table 2.). To ensure model reliability and interpretability, key implementation details are provided. The 2,276 spatial units were split into training (70%), validation (10%), and test (20%) sets. Hyperparameters were optimized using a two-step strategy: an initial grid search to define core parameter ranges, followed by Bayesian optimization fine-tuned with validation RMSE as the early‑stopping criterion. To prevent overfitting, we employed early stopping, L1/L2 regularization (reg_alpha=1.06, reg_lambda=5.00), and stochastic sampling of data (subsample=0.63) and features (colsample_bytree=0.52). The model generalizes well, with test RMSE (0.0251) close to training RMSE (0.0217), indicating no significant overfitting and reliable predictive performance.

Table 2.

Optimal parameters of the XGBoost model.

Parameter	Value	Parameter	Value
n_estimators	1053	colsample_bytree	0.52
max_depth	8	Cross-validation	5
learning_rate	0.01		0.66
reg_alpha	1.06	RMSE	0.0251
reg_lambda	5.00	MSE	0.0006
Subsample	0.63	MAE	0.0172

XGBoost offers superior accuracy over linear models but suffers from low interpretability—the “black-box” problem. This study thus uses the SHAP framework²⁸ to resolve this. Based on Shapley values, SHAP quantifies the marginal contribution of each feature, making model decisions interpretable. The formula for computing the predicted value of a sample is as follows:

12

Here, represents the average predicted value across all samples, denotes the value of sample for feature , and is the total number of features in the sample. The method quantifies the contribution of each feature based on the machine learning model’s predictions, thereby addressing the black-box problem by providing interpretable insights into model behavior.

Result

Spatial distribution characteristics of single vitality

Social vitality

Social vitality in Jinan exhibits spatial clustering along major transport corridors (Fig. 4). The highest values (0.39–1.0) concentrate in the Quancheng Road–Baotu Spring–Daming Lake area, which serves as the city’s historic, cultural, and commercial core. Secondary centers, such as Hanyu Jingu, Hongjia Square, Jinan West Railway Station, and Shandong Provincial Sports Center, show moderate vitality (0.20–0.39), reflecting the outward diffusion of urban functions clustered around transit hubs and public facilities, demonstrating a transit-oriented development pattern. Vitality extends eastward along Jingshi Road toward sub-centers, while other key corridors like Luoyuan Avenue, Jiefang Road, Gongye South Road and Weier Road, Yingxiongshan Road further channel activity. By contrast, northern Tianqiao, western Huaiyin, and eastern Licheng display low vitality, functioning mainly as residential “dormitory towns” with limited services. Industrial zones and ecological areas such as Mount Qianfo and Xiaoqing River also lack urban functions. Overall, a “core–corridor–multi-node” structure emerges with radial decay. Development favors eastern and southern expansion, aligning with Jinan’s spatial strategy, while northern and western progress remains limited.

Fig. 4.

Spatial distribution (left) and cold and hot spot analysis (right) of social vitality.

Economic vitality

In terms of economic vitality, a marked disparity is evident: the high values ranging from 0.46 to 1.0 are confined solely to the traditional central district along Quancheng Road-Luoyuan Avenue and the newly developed central business district along Jingshi East Road, which makes up a unipolar structure, as shown in Fig. 5. Hotspot analysis is used to detect clustering patterns, revealing significant clustering phenomena with a significance level of 99%. On the other hand, large areas in the north, west, and east constitute cold spot regions, with reliability ranging from 90% to 99%, representing economic vitality depressions. This core-periphery pattern, with a strong siphon effect emanating from the core, goes up against the peripheral area’s minimal absorption capacity due to their fragile industrial bases and poor accessibility. This further strengthens the city’s eastward development, forming an economic corridor with vitality extending eastward along Jingshi Road.

Fig. 5.

Spatial distribution (left) and cold and hot spot analysis (right) of economic vitality.

Cultural vitality

Unlike economic or social vitality, cultural vitality forms a multi-nucleated network (Fig. 6), with high-value areas in historic cores and modern educational zones along Jingshi Road. Its 99% confidence hotspots are spatially fragmented, exhibiting weaker polarization and more limited diffusion. Driven by both historical and modern facilities, the pattern is more balanced overall, though cultural services remain scarce in peripheral regions.

Fig. 6.

Spatial distribution (left) and cold and hot spot analysis (right) of cultural vitality.

Comprehensive dynamic spatial coupling characteristics

Comprehensive vitality composition analysis

The Comprehensive Vitality (COV) was constructed using the CRITIC method, which assigns weights by balancing intra-indicator variability and inter-indicator conflict. As Table 3. shows, SV has the highest spatial variability (0.0937), while EV shows the strongest conflict (-0.7390), indicating its inverse correlation with other dimensions. Figure 7 reveals that CV holds the strongest synergy with EV (correlation is 0.45). Consequently, CV received the highest weight (40.9%) in the COV, followed by SV (36.7%) and EV (22.4%). This reflects a “culturally led, socially supported, and economically complementary” mechanism, where CV’s high synergy makes it the core driver, SV’s high variability ensures broad representation, and EV’s high conflict suppresses its weight.

Table 3.

CRITIC index.

Vitality type	Social vitality (SV)	Economic vitality (EV)	Cultural vitality (CV)
Standard deviation	0.0937	0.0720	0.0435
Conflict index	− 0.5098	− 0.7390	− 0.6700
Information amount	− 0.0478	− 0.0531	− 0.0291

Fig. 7.

Single vitality correlation and weight.

Comprehensive vitality spatial distribution

The spatial pattern of comprehensive urban vitality arises through the interpenetration of social, economic, and cultural dimensions (Fig. 8). Areas of high value are found centered on Quancheng Road and the eastern CBD, stretching along Jingshi Road, the primary development axis of the city. This core is economic, socially and culturally functional, reflecting synergistic active functions through functional mixing. Confirmed through hotspot analysis, a statistically significant cluster—at a 99% confidence level—is found around this core; large northwestern, western, and northeastern peripheral cold spots form an overall pattern of internal disparity. The results of the spatial autocorrelation analysis (Table 4) indicate that the Moran’s I value for comprehensive vitality reaches 0.4214 (Z-score = 133.1396). Additionally, all individual vitality dimensions exhibit significant positive spatial autocorrelation, further validating the statistical reliability of the aforementioned agglomeration characteristics. Overall, this indicates that comprehensive vitality exhibits a concentric decline pattern radiating outward from the core area, yet shows axial expansion along the city’s main thoroughfare (Jingshi Road), demonstrating that infrastructure significantly influences the distribution of urban vitality. The overall structure of COV in Jinan’s central urban area exhibits characteristics of “high polarization, core dominance, and axial extension,” with development in the southern and eastern directions far outpacing that in the west and north—once again highlighting the current state of development imbalance.

Fig. 8.

Spatial distribution (left) and cold and hot spot analysis (right) of comprehensive vitality.

Table 4.

Spatial autocorrelation analysis of vitality.

Vitality type	Social vitality (SV)	Economic vitality (EV)	Cultural vitality (CV)	Comprehensive vitality (COV)
Moran’s I	0.3616	0.2685	0.2067	0.4214
Z score	113.0259	88.0312	49.2232	133.1396

SV, EV, CV and COV all passed the significance level test (p < 0.01).

Nonlinear impact of built environment on urban comprehensive vitality and threshold effect

The characteristic importance of the built environment

Figure 9 shows a series of SHAP feature importance and bees warm plots for the built environmental factors in terms of their impact on urban vitality. Here again, objective functional indicators prevail: the leading indicator is POI blending degree (mean SHAP: 0.24), followed immediately by POI density (0.18) and spatial integration (0.16). This once again highlights that functional diversity holds absolute primacy in predicting urban vitality, far surpassing the importance of density or geometric accessibility. Traditional spatial indicators (integration and openness) retain their reference value, but their influence coefficients have further declined to 0.16 and 0.10. Subjective perception indicators such as green space visibility rate, walkability, and image property exhibit relatively weaker explanatory power for Jinan’s urban vitality (SHAP values below 0.05). This indicates that in the formation of urban vitality, the role of functional supply significantly outweighs human-scale perceptions. SHAP dependence plots reveal unique nonlinear patterns. POI blending degree shows a suppression-promotion threshold around 1.5, consistent with Yue²⁹. Positive but diminishing returns are seen for POI density over 4,000 units/km². Integration peaks at 0.4–0.8, where optimal accessibility is found. Openness demonstrates an inverted-U effect, being beneficial when below 0.2 but inhibitory when above 0.3. Other perceptual and accessibility metrics have weak or no marginal effects, further confirming the secondary role of experiential factors.

Fig. 9.

Built environment feature importance and bees warm plot.

Interaction and threshold effects of main built environment indicators

This study explores threshold behaviors and interactions among the five most significant built environment features identified by XGBoost-SHAP (Fig. 10). The POI blending degree and POI density exhibit the widest range of SHAP values, confirming their role as primary drivers of urban vitality. A steep threshold exists for the POI blending degree at 1.5, below which it inhibits vitality and above which it strongly promotes vitality, particularly in areas with high spatial integration (SHAP > 0.2). This critical threshold (> 1.5) was determined through data-driven identification from the SHAP dependence plot, validated by statistically significant breakpoint analysis (p < 0.01, Chow test), this reflects a clear synergy between functional diversity and spatial connectivity. The effect of POI density shows a pattern of diminishing marginal returns. SHAP values plateau after approximately 4000 facilities/km², with the marginal effect (ΔSHAP/Δdensity) decreasing substantially(from 0.00012 to 0.00003). The threshold of below 4000/km² is supported by the 95% confidence interval of POI density in the top 10% high-vitality blocks. Spatial integration exhibits an optimal range for enhancing vitality between 0.4 and 0.8; exceeding 0.8 may exacerbate congestion and diminish positive effects. This optimal range (0.4–0.8) was determined from the peak of the SHAP dependence plot and aligns with established optimal integration ranges for urban blocks in comparable contexts⁸.

Fig. 10.

SHAP dependence plots for the five most influential built environment features.

Furthermore, the openness ratio follows an inverted U-shaped relationship with vitality, promoting it at moderate levels (around 0.2) but becoming inhibitive at higher levels (> 0.3), especially in areas with low integration. This positive effect is most pronounced in high-density zones (> 10,000 persons/km²) near subway stations. Key interactions were identified: a high POI blending degree (> 1.5) combined with moderate integration (0.4–0.8) yields high synergistic effects (SHAP > 0.45), as observed in Jinan’s CBD. Population density also moderates the role of openness, where moderate openness (0.15–0.25) promotes vitality in low-density areas but reduces it in high-density areas. Thus, while POI density and openness jointly influence vitality at moderate levels, openness loses its regulatory role at high values.

In summary, this study advances urban vitality theory by showing that urban vitality operates as a complex adaptive system, governed by nonlinear thresholds and context-dependent interactions among built environment factors. By identifying precise parameters, we shift the focus from linear, single-factor approaches toward the synergistic optimization of multiple elements within bounded effective ranges. The analysis further reveals that the role of one factor can change based on the conditions of another; for example, openness promotes vitality in low-density areas but may inhibit it in high-density settings. This underscores a layered logic in vitality formation: objective functional-spatial factors form a necessary foundation, especially in cities like Jinan that are still developing core urban functions. Only once this foundation is adequately established can perceptual refinements yield significant returns. Ultimately, this work offer a theoretically grounded, parameter-informed framework for targeted and adaptive urban planning, supporting a transition from static “element-based” thinking toward dynamic “system-relationship” and “precision-intervention” approaches.

Planning strategy

Based on the spatial pattern of “east-strong west-weak, south-rising north-stagnant” and the identified nonlinear threshold effects of the built environment in Jinan’s central urban area, this study proposes targeted vitality-enhancement strategies. To ensure scientific rigor, the study area is first classified into four strategic zones (Fig. 11) according to spatial-functional attributes, vitality characteristics, and key built environment thresholds, aligning with the city’s territorial plan (2021–2035) and the 2,276 analyzed blocks. Tailored planning interventions are then outlined for each zone, moving from core-led spillover to functional supplementation in peripheral areas.

Fig. 11.

Urban strategic zoning.

The Eastern Core Area (bounded by key roads such as Jingshi East Road and Fenghuang Road) serves as the modern commercial-service core, characterized by high POI blending (> 1.5), density (3800–4200 units/km²), and top-tier comprehensive vitality. Here, strategy focuses on sustaining and radiating its leading role. Within a 3-km core, deploying micro-commercial complexes and community cultural stations can maintain a POI blending degree above the promotive threshold of 1.5. Pedestrian corridors along Jingshi Road and Luoyuan Street should be enhanced to support walkability. Specifically within its CBD sub-zone, built environment parameters should be finely tuned: the proportion of commercial/office POIs controlled below 40% (as observed in top-vitality blocks), POI blending kept at an optimal 1.8–2.2, POI density capped at 4,000 units/km² to avoid diminishing returns, spatial integration optimized to 0.6–0.8, and openness maintained at 0.2–0.25 to leverage its inverted U-shaped benefit.

The Historic Districts (centered on Daming Lake-Baotu Spring) exhibit “high perception + medium function” traits, with cultural facility density of 8–10 units/km² and openness between 0.15 and 0.2. Strategies here prioritize perceptual quality and cultural vitality: preserving the optimal openness range (0.15–0.2) to balance comfort and interaction, enhancing walkability through streetscape improvements like tree planting, and actively increasing cultural facilities to match the density benchmark of high-vitality historic blocks.

The Western High-Speed Rail Area (around Jinan West Railway Station) currently shows medium social but weak economic-cultural vitality. Development should prioritize functional supplementation and transportation synergy, focusing on cultivating a commercial-cultural tourism district around the transportation hub and revitalizing industrial heritage corridors. Increasing bus stop density and establishing integrated “bus hub + community service” nodes along major thoroughfares are key to improving accessibility.

The Peripheral Residential Areas (e.g., northern Tianqiao) function as “dormitory towns” with low POI density (1,200–2,500 units/km²) and vitality. The primary strategy is to elevate basic functional supply and accessibility: achieving a POI integration index above 1.0, developing community centers, optimizing public transit coverage with micro-circulation bus routes, and gradually increasing POI density toward the identified effectiveness threshold.

To implement these spatially differentiated strategies, cross-sectoral policy coordination is essential. This includes employing big data for dynamic monitoring, establishing dedicated vitality-enhancement funds, and offering targeted incentives to businesses to foster synergistic cultural, economic, and social benefits.

Conclusion and discussion

Conclusion

This study systematically investigates the nonlinear impacts of the built environment on urban vitality in Jinan’s historic-modern urban context using multi-source data and explainable AI. The principal findings are: (1) The individual vitality dimension exhibits a distinct spatial pattern, while comprehensive vitality reveals a mechanism characterized by “cultural leadership, social support, and economic supplementation.” Social vitality is concentrated in the Quancheng Road-Baotuquan-Daming Lake area and expands outward along the Jingshi Road axis; economic vitality shows bipolar aggregation in the eastern CBD; cultural vitality demonstrates a multi-centered distribution pattern. In constructing comprehensive vitality, cultural vitality holds the highest weight (40.9%) due to its significant synergistic effect with economic vitality (r = 0.45), followed by social vitality (36.7%) and economic vitality (22.4%). The pronounced spatial clustering of comprehensive vitality (Moran’s I = 0.4214, p < 0.01) aligns with Jinan’s development pattern of “strong east, weak west.”(2) SHAP analysis reveals that POI blending degree (0.24), density (0.18), and spatial integration (0.16) are the main drivers, collectively explaining over 60% of vitality variation. Meanwhile, subjective perceptual indicators have little effect (SHAP < 0.05), showing that Jinan’s vitality gaps are due more to functional supply than to human-scale perceptions. (3) Key thresholds emerge: POI blending degree activates beyond 1.5; POI density shows diminishing returns above 4000 /km²; Integration is most effective at 0.4–0.8. Openness has an inverted U-shaped effect (< 0.2 promotes, > 0.3 inhibits). The strongest synergy (SHAP > 0.45) comes from high POI blending (> 1.5) paired with moderate integration (0.4–0.8). A clear contrast exists: historic districts have high comfort but low density, while modern CBDs have high density but low openness. This implies that historical areas need moderated enclosure, modern zones need better functional blending, and peripheral areas need more POIs and transport upgrades.

This research is not only applicable to Jinan, but also provides a reference for other similar cities that possess both historical and modern textures. The research enriches the multi-dimensional driving system of urban vitality, supplements the 5D framework with threshold effects, and promotes the transformation of urban vitality research from “linear correlation” to “nonlinear threshold governance”.

Discussion

This study is based on the empirical evidence of Jinan City, promoting the development of the theory of urban vitality, and at the same time connecting methodological innovation with theoretical development. Firstly, it expands Jacobs’ emphasis on diversity and pedestrian flow by determining the synergy effect of “cultural dominance, social support, and economic supplementation”, where the weight of cultural vitality is 40.9% and the correlation coefficient with economic vitality is 0.45, serving as a binding force between social and economic activities. This challenges the traditional dominance of economic factors³⁰ and supplements the cultural dimension that is often overlooked in the traditional framework. Secondly, by integrating 12 building environment indicators (including perception indicators) into a unified analytical framework, this study refined Ewing’s 5D model. The results show that objective functional factors (poi blending degree, density and spatial integration) explain over 60% of the vitality changes, while the perceptual indicators have a limited impact in the context of Jinan (SHAP < 0.05), clarifying the “functional supply priority” principle of cities where history and modernity intermingle. Thirdly, by determining the optimal integration range of vitality (0.4–0.8), the spatial syntactic theory is enriched. Beyond this range, the benefits will decrease - this finding challenges the linear assumption and reveals the diminishing marginal returns of spatial accessibility. Fourth, the detection of key thresholds (for example, POI blending degree > 1.5, density ≤ 4000 /km²) transforms qualitative nonlinear descriptions into actionable planning rules, and proposes a “threshold based” vitality enhancement theory, solving the problem of overemphasizing a single indicator in the past³¹.

The weak role of the perception indicators in Jinan City, which is context-dependent, can be attributed to the following background factors: (1) The current development stage of the city prioritizes functional supplementation, especially in the peripheral areas where POI and traffic density are below the optimal threshold; (2) There is a certain spatial offset between perception and function. Historical districts have a higher level of perceptual comfort and a lower functional density, while modern CBDS present the opposite pattern, offsetting the perceptual effect. (3) Jinan’s vitality has long relied on functional agglomeration. In the top-level vibrant districts, the degree of POI mixing and integration is relatively high, but the perceived score is not high. These findings also apply to Chinese transitional cities where historical and modern structures are mixed and functional supplies are unbalanced. These cities should give priority to functional supplementation. They are not very suitable for mature cities with developed infrastructure and high-quality perceived environments, where the role of perception is more important.

The limitations of this study include the use of data from different times, particularly the 2020 population data and the 2023–2024 and POI data, as well as sparse street view sampling from surrounding areas, which may affect perception metrics. Meanwhile, the stability of the threshold conclusions drawn in this study in different scenarios remains to be verified. Future research should integrate dynamic population estimation, establish a perception-behavior-vitality path model by combining survey and behavioral data, and verify the proposed thresholds through field studies and policy pilots.

Author contributions

Mingyang Yu and Qingrui Ji wrote the initial draft and made revisions to the manuscript together; Xiangyu Zheng is responsible for data acquisition, data visualization, formula analysis; while Weikang Cui is responsible for the results verification. All authors reviewed the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China(grant number: 41801308) and Shandong Provincial Department of Housing and Urban-Rural Development Project (grant number:2025KYKF-CZJS220).

Data availability

The Street View Image datasets used and analyzed during the current study are available from the corresponding author on reasonable request. The link to the street view image dataset used in this research is: https://www.scidb.cn/s/7ZBnM3.

Declarations

Competing interests

The authors declare no competing interests.