Comprehensive walkability assessment of urban pedestrian environments using big data and deep learning techniques

Abstract

Assessing street walkability is a critical agenda in urban planning and multidisciplinary research, as it facilitates public health, community cohesion, and urban sustainability. Existing evaluation systems primarily focus on objective measurements, often neglecting subjective assessments and the diverse walking needs influenced by different urban spatial elements. This study addresses these gaps by constructing a comprehensive evaluation framework that integrates both subjective and objective dimensions, combining three neighbourhood indicators: Macro-Scale Index, Micro-Scale Index, and Street Walking Preferences Index. A normalization weighting method synthesizes these indicators into a comprehensive index. We applied this framework to assess the street environment within Beijing’s Fifth Ring Road. The empirical results demonstrate that: (1) The framework reliably reflects the distribution of walkability. (2) The three indicators show both similarities and differences, underscoring the need to consider the distinct roles of community and street-level elements and the interaction between subjective and objective dimensions. (3) In high-density cities with ring-road development patterns, the Macro-Scale Index closely aligns with the Comprehensive Index, demonstrating its accuracy in reflecting walkability. The proposed framework and findings offer new insights for street walkability research and theoretical support for developing more inclusive, sustainable and walkable cities.

Introduction

Walking is one of the healthiest mode of travel and the most common form of recreational physical activity, promoting public health, strengthening social interactions, and yielding significant socio-economic benefits^1,2,3. To foster a conducive walking environment, urban planners and policymakers increasingly prioritize a “pedestrian-first” principle in urban planning and design^4,5.

Extensive literature explores how urban spatial characteristics influence street walkability. Research can be categorized into two types: objective measurements, which examine community-level elements (e.g., amenities, street connectivity) and street-level factors (e.g., greenery, vehicular traffic)⁶, and subjective measurements, which capture people’s perceptions of the street environment through environmental audits, surveys, and other methods. While both systems are effective⁷, many scholars argue that perceived walkability often has a greater impact on walking comfort than direct environmental conditions^8,9. Due to the challenges of obtaining micro-level data and conducting surveys¹⁰, most studies focus solely on objective measurements, lacking comprehensive assessments¹¹. Some studies consider both dimensions but are limited in scope, providing insights not applicable to broader urban settings^12,13.

According to the theory of hierarchical walking needs, community and street-level elements impact various levels of walking needs differently. Community-level elements address lower-level needs like connectivity and basic services, while street-level elements relate to higher-level needs such as safety and comfort¹⁴. Most walkability studies overlook these differences^15,16, with only a few considering their impact^6,12.

The advancement of urban big data and computer vision technology has significantly enhanced street walkability research in terms of data sources, accuracy, and methodologies. From low-resolution GIS data to current GPS and mobile signalling data, and from manual audits to virtual and automated audits using computer algorithms, walkability research methods have evolved, offering new perspectives for walking environment assessments¹⁷.

Based on the identified research gaps and technical considerations, the primary contributions of this study are as follows:

• 1. This study introduces a novel street walkability evaluation framework that integrates both subjective and objective dimensions. By leveraging big data and deep learning technologies, the framework efficiently assesses urban walking environments at both the community and street levels. Its flexibility allows for application in walkability studies across various complex urban settings.

• 2. Through quantitative analysis, this research reveals the interactive characteristics of three key street indicators—the Macro-Scale Index, Micro-Scale Index, and Street Walking Preferences Index. The study also explores the underlying causes behind these interactions, offering theoretical support for future urban planning and street design.

• 3. The empirical results demonstrate that Macro-Scale Index more accurately reflect street walkability in high-density cities characterized by ring-road development patterns. This study uncovers the complex relationship between urban morphology and walkability, offering new perspectives for optimizing walking environments under various urban development pattern.

Literature review

Paradigm melding deep learning algorithms and streetscape images

Early built environment studies often relied on manual data collection through field visits, limiting the scale and precision of research. Advances in sensing and computer technologies have introduced a paradigm combining deep learning algorithms with streetscape images, providing a reliable method for obtaining micro-level street data and significantly enhancing the ability to describe built environment characteristics^6,18,19.

Some studies have focused on validating image data against actual data. For instance, Yin et al. used image recognition algorithms to detect pedestrian numbers in streetscape images from New York State and validated these counts against over 200 street segments²⁰. Zhang et al. demonstrated how streetscape images and taxi data could train predictive models, comparing the predicted 24-hour taxi flow variations on 917 streets in Beijing with actual data, proving high accuracy and stability²¹.

Other studies have assessed street walkability using streetscape images. Rui utilized the DeepLab model for pixel-level segmentation of Google Street View and self-collected images, comparing greenery, openness, walkability, and imageability dimensions to establish an effective evaluation system. Researchers have also used street indicators from images to explore correlations with travel modes, route choices, and physiological health²². For example, Koo et al. used personal attributes as control variables to study the impact of street greenery on travel modes⁶, while Wang et al. found that a higher sky view factor positively impacts the mental health of senior residents²³.

This research paradigm offers three main advantages: (1) Streetscape images are easily accessible and include elements like trees and cars, often missed by satellite images^21,24,25. (2) It saves significant time and labour compared to manual audits, while achieving good data accuracy²⁶. (3) It provides a diverse perspective to evaluate subjective perceptions and human behaviour on streets²⁵.

Objective evaluation systems for street environments

Objective evaluation systems for street walkability are multi-attribute frameworks that include indicator selection, variable processing, and weighting and aggregation²⁷. Nearly 200 different indicators have been used in urban street walking environment studies^28,29. Many focus on community-level elements, with the most typical being the “D” indicators. Robert Cervero et al. proposed the 3D model—Density, Diversity, and Design—in 1997, noting these indicators’ significant influence on travel frequency and mode³⁰. Ewing expanded this to a 5D model by adding destination accessibility and distance to transit³¹, and further to a 7D model by including demand management and demographic attributes³².

These “D” indicator models have been widely adopted and expanded due to their universality and flexibility. Notable examples include the Walkability Index (WAI) and Walk Score (WS)^33,34. The development of data science has led to more systematic studies on built environment and pedestrian activities, expanding the focus to street-level factors and personal and socio-economic attributes^{22,35,36,37,38,39}.

After determining appropriate evaluation factors, converting qualitative indicators into quantitative ones and using reasonable methods to measure their weights is crucial. Many studies emphasize using subjective and objective weighting methods^27,40. Among subjective methods, the Analytic Hierarchy Process (AHP) and expert scoring are widely used but are often arbitrary and influenced by personal biases^41,42. Objective methods like entropy weighting and principal component analysis (PCA) reduce subjective bias but are sensitive to outliers and data quality^{13,43,44,45,46}. To reduce the potential errors, this study utilizes the stated preference (SP) method and discrete choice model (DCM) for weighting factors. This approach enables efficient data collection while accurately capturing pedestrian preferences and behavior. By combining qualitative and quantitative analyses, this methodology accommodates diverse data types and effectively addresses data quality issues, offering a flexible solution for robust pedestrian perception analysis^47,48,49.

Subjective perceptions in walkability measurement

Although many paradigms have become proficient in data acquisition and processing, they often neglect pedestrians’ subjective perceptions. Researchers typically use manual audits, virtual audits, and automated audits to capture people’s intuitive feelings⁵⁰.

Manual audits, using questionnaires, interviews, or field surveys, accurately express individual subjective perceptions but are costly and unsuitable for large-scale assessments^7,51,52,53. Virtual audits based on streetscape images are simpler, more efficient, and cost-effective but cannot observe finer street characteristics and qualitative features^54,55,56. Automated audits using streetscape images and machine learning algorithms are more efficient, consistent with virtual audits, and suitable for large-scale assessments^57,58. Currently, some studies employ semantic segmentation models alongside machine learning algorithms to calculate urban perception scores^59,60. However, these methods fail to extract the topological features of street spaces, thereby limiting their ability to capture robust nonlinear relationships between street environments and human perception⁶¹. Consequently, this research adopts an end-to-end convolutional neural network model to directly extract high-dimensional features from street view images for predicting urban perception scores.

Research methods and data

Research framework

The evaluation framework we developed (Fig. 1) involves three main steps.

Step 1: Calculation indicators.

Data preparation and preprocessing: Collect and clean necessary data.

Construct models: Calculate the Macro-Scale Index (MAI), Micro-Scale Index (MII), and Street Walking Preferences Index (SPI) using different models.

Step 2: Statistical analysis.

Use mathematical statistical methods to analyse the correlation between streetscape elements and the three indices, and examine the similarities and differences among the indices. This elucidates the contribution of different elements and indices to the street walking environment.

Step 3: Comprehensive index calculation.

Calculate the Comprehensive Index (CI) using a normalization weighting formula based on walkability differences among neighbourhoods. Map the comprehensive evaluation results onto a land use classification map for further analysis.

Fig. 1

The research framework diagram.

Objective measurement methods

Calculating the macro-scale index

As discussed in the literature review, “D” models are widely applicable and have been adapted for various local studies^62,63,64. To represent the impact of community-level elements on the walking environment, this study introduces the Walk Score, reflecting the level of accessibility to walkable destinations within a certain range around the street, commonly used as a street walkability indicator. Drawing upon the calculation methodology used by Walk Score, the process of calculating the MAI involves three distinct phases: configuring the facility classification table and weights, calculating the basic MAI, and correcting the basic MAI³⁸.

First, we adopt the facility classification standards and weighting system of Walk Score, while referencing the daily facility allocation standards outlined in China’s Urban Residential Area Design Code⁶⁵. We classified and localized certain facilities, ultimately selecting five primary facility types. Corresponding weights were assigned based on the frequency of daily use of these facilities (Table 1)⁶³.

Table 1 Facility classification and weight table.

Second, In real-world walking scenarios, the distance decay effect refers to the phenomenon whereby the accessibility and convenience ratings of a given facility progressively diminish as the distance increases. In other words, people generally prefer to walk to nearby facilities or destinations, and their willingness to walk decreases correspondingly with greater distances. The Walk Score methodology defines three decay thresholds: within the comfortable walking distance (400 m–5 min), no decay is applied; within the tolerable walking distance (400 to 1,600 m or 5 to 20 min), a decay of 12% is applied; beyond the resistant walking distance (2,400 m–30 min), a decay of 100% is applied⁶⁶. Based on the walking characteristics reported in the census of residents within our study area, we have adjusted these three decay thresholds to 480 m (6 min), 1,520 m (19 min), and 2,080 m (26 min), respectively⁶⁷. Using a Gaussian function, we fit the distance decay thresholds to obtain the walking decay curve (Fig. 2). Based on the initial weight ratio table and the walking decay curve, we calculate the basic MAI for each streetscape image collection point, as shown in Eq. (1):

$$\:\begin{array}{c}{W}_{bj}\text{\:}={\sum\:}_{i=1}^{n}{W}_{i}\times\:{(1-D}_{ij})\end{array}$$

(1)

W_bj represents the basic MAI of the j-th collection point, W_i is the impact weight of the i-th type of facility, and n represents the total number of facility categories at the j-th collection point. D_bj refers to the decay coefficient corresponding to the actual distance between the i-th facility type and the j-th collection point.

Third, in a specific area, a higher number of road intersections per unit area and shorter block lengths indicate a more dense road network. This increased density provides pedestrians with more route options and facilitates easier access to destinations⁶³. Consequently, we used two quantitative data—road intersection density and block length (Table 2)—to correct the basic MAI for each point and obtain the final single-point MAI.

$$\:\begin{array}{c}{W}_{j}={W}_{bj}\times\:\left(\frac{100-{d}_{cross}}{100}\right)\times\:\left(\frac{100-{d}_{length}}{100}\right)\end{array}$$

(2)

In the formula, W_j and W_bj represent the final single-point MAI and the basic MAI of the j-th collection point, respectively. d_cross and d_length represent the decay rates for intersection density and block length, respectively.

Fig. 2

Walking decay curve in time.

Table 2 Intersection density and block length decay rate comparison.

Stated preference experiment

We used stated preference (SP) experiments to collect behavioral preferences regarding streets. SP experiments present respondents with virtual scenarios to express their preferences, offering advantages such as controllable variable quantities and correlations, aligning data with theoretical expectations. The experiment process consists of four steps⁶⁸:

• 1. 1. Attribute selection: Not all features within the built street environment influence pedestrians’ walking choices²⁸. Based on research on the built environment—and considering multicollinearity among attributes^29,69 as well as the optimal number of attributes for SP experiments^70,71—we selected seven micro-elements most relevant to pedestrian perceptions as indicators of walking preferences. All seven attributes are categorical variables coded using dummy variables (Table 3).

• 2. 2. Scenario selection: To avoid respondent fatigue, which affects data accuracy, we set the number of scenario selections to 12⁷². In each selection, respondents face three distinct street scenarios (Fig. 3) and choose one. Each option is unlabeled and presented as an image to avoid assumptions and enhance authenticity^73,74.

• 3. 3. Experiment design: The key point in SP experiment design is determining which combinations of attribute levels to present and how to allocate these scenarios in the choice set. Many hypothetical scenario experiments use orthogonal design⁶⁸, beneficial for linear regression models. However, we used a nonlinear model, so we adopted the more flexible and predictive D-efficient design^75,76.

• 4. 4. Sample size determination: Existing research mainly uses three methods: literature reference, empirical estimation, and parameter calculation. To improve experiment quality and statistical power, we adopted the parameter method proposed by De Bekker-Grob EW to calculate the sample size⁷². The calculation formula is as follows:

$$\:\begin{array}{c}N>{\left(\left({z}_{1-\beta\:}+{z}_{1-\alpha\:}\right)\sqrt{\sum\:_{\gamma\:k}\:}/\delta\:\right)}^{2}\end{array}$$

(3)

In the formula, N is the minimum required sample size, and the final collected sample size should be greater than the sample size required for each parameter. We set the significance level α to 0.05 and the statistical power 1−β to 80% to ensure that we can detect differences of practical significance. According to these parameters, the corresponding values from the standard normal distribution table are z_1−β = 0.84 and z_1−α = 1.96. k is the number of parameters to be estimated, which is 12 in this study; $\:\sum\:_{\gamma\:k}\:$represents the k-th diagonal element in the asymptotic variance-covariance matrix (AVC). The utility value δ is obtained from a pilot survey of 40 people (Table 3).

Table 3 Attribute names and levels.

Fig. 3

Street scenarios test demonstration.

Discrete choice model

This study employs the conditional Logit model to quantitatively analyze pedestrians’ rational decision-making processes on streets⁷⁷. While the general Logit model includes personal characteristic variables, the conditional Logit model focuses on the impact of attribute changes on respondents’ choices. Since we primarily explore the impact of street environments on pedestrian walking without considering personal attributes, the conditional Logit model is the most appropriate discrete choice model. The theoretical basis of the conditional Logit model is random utility theory⁷⁸. To simulate the psychological activities of pedestrians, this theory assigns an attractiveness or utility value to each choice set. When pedestrian $\:i$ faces $\:M$ street choices, their preference for a particular street $\:j\:$can be described by the utility value $\:{T}_{ij}$. $\:{T}_{ij}$ is an unknown function, but for convenience in calculation, it is often assumed to be a linear function, described by fixed utility $\:{V}_{ij}$ and random utility $\:{\epsilon\:}_{ij}$. The fixed utility is represented by the combination of street environment characteristics $\:{X}_{j}$, while the random utility includes unobserved utility and random error. Therefore, $\:{T}_{ij}$​ can be expressed as:

$$\begin{gathered} T_{{ij}} \;\left| { = V_{{ij}} + \varepsilon _{{ij}} } \right. \hfill \\ \;\;\;\;\left| { = \beta _{{ij}} X_{{ij}} + \varepsilon _{{ij}} } \right. \hfill \\ \end{gathered}$$

(4)

In this equation, $\:{\beta\:}_{ij}$ represents the vector of parameters to be estimated for the street attributes of the j-th street option, which is calculated from statistical data. Furthermore, based on the basic assumption of economics, decision-makers always seek to maximize utility. That is, when pedestrian $\:i\:$chooses street $\:j$, we have $\:{T}_{ij}>{T}_{im}(m\in\:M,m\ne\:j)$. Therefore, the probability that pedestrian $\:i\:$chooses street $\:j$ is given by:

$$P_{i} \left( j \right) = P\left( {T_{{ij}} > T_{{im}} } \right),\left( {m \in M,m \ne j} \right)$$

(5)

Substituting Eq. (4) into Eq. (5), and assuming that $\:{\epsilon\:}_{ij}\:$follows a generalized extreme value distribution, we can derive the form of the conditional Logit model. Thus, the probability that pedestrian $\:i\:$chooses street $\:j$ is given by:

$$\:\begin{array}{c}{P}_{i}\left(j\right)=\frac{\text{exp}\left({\beta\:}_{ij}{X}_{im}\right)}{{\sum\:}_{m=1}^{M}\text{exp}\left({\beta\:}_{ij}{X}_{im}\right)}\end{array}$$

(6)

Using Eq. (6), we fit the survey data and estimate parameters using the maximum likelihood method. The estimated parameters reflect the impact of different street attribute levels on choice probability and reveal pedestrians’ subjective preferences for streets with various element combinations. This further indicates the street’s walkability.

Micro data acquisition and perception score assessment

Semantic segmentation and instance segmentation models

To quantify the impact of street micro-elements on pedestrians, we use semantic segmentation and instance segmentation models based on deep learning frameworks to obtain quantitative indicators for seven street elements. Greenery visibility, interface transparency, relative sidewalk width, and barrier separation are calculated based on pixel proportions in the images, while the number of street facilities, motor vehicles, and pedestrians are derived from detection results (Table 4).

We used a semantic segmentation model based on the Pyramid Scene Parsing Network (PSPNet) and the Deep Residual Network (ResNet-101). This combination leverages the deep feature extraction capabilities of ResNet-101 and PSPNet’s advantages in handling multi-scale information⁷⁹. The model was pre-trained using the Cityscapes dataset, which provides 19 categories and is well-suited for semantic segmentation tasks in urban scenes⁸⁰. For instance segmentation, we used the Mask R-CNN framework with an improved ResNet-50 backbone network, specifically designed for instance segmentation tasks (Fig. 4). Numerous studies have demonstrated the excellent performance of these deep learning models in recognizing urban street images^81,82,83.

Table 4 Attribute names and level classification.

Fig. 4

Semantic segmentation process and the perceptual scoring.

Subjective perception model

This study utilizes the Place Pulse 2.0 (PP2.0) dataset, an open-access large-scale streetscape image database from the Massachusetts Institute of Technology. The dataset includes 110,000 images from over 50 regions worldwide, aimed at studying the perceived quality of urban environments through crowdsourcing. Participants compared randomly selected image pairs based on one of six dimensions: “beautiful,” “wealthy,” “lively,” “safe,” “boring,” and “depressing,” with three options: the left image, the right image, or both images are the same (Fig. 5). From 2013 to 2016, 1.17 million paired comparison data points were collected. Using the Trueskill algorithm, these comparisons were converted into perception scores across the six dimensions^86,87.

The model training process is shown in (Fig. 5). We treat human perception prediction as a binary classification task, as classification tasks offer higher stability in emotion prediction compared to regression tasks, reducing uncertainty from human subjective cognition^88,89.

First, to classify positive and negative samples, we calculated the mean and variance of the scores for each dimension, using these to determine thresholds for classifying images (Formula 7):

$$\:\begin{array}{c}{E}_{i,d}=\left\{\begin{array}{c}0,\:\:\:{\:Q}_{i,d}\le\:{\stackrel{-}{Q}}_{d}-{\sigma\:}_{d}\\\:1,\:\:\:{Q}_{i,d}\ge\:{\stackrel{-}{Q}}_{d}+{\sigma\:}_{d}\end{array}\right.,\:\:\:d\in\:\left\{beautiful,\:wealthy,\:lively,\:safety,\:boring,\:depressing\right\}\end{array}$$

(7)

In the formula,$\:{\stackrel{-}{Q}}_{d}$ represents the average score of all samples in a particular dimension, and $\:{\sigma\:}_{d}$ represents the standard deviation. $\:{Q}_{i,d}$ indicates the score of image i in this dimension. $\:{E}_{i}$ is set to 0 if image i is a negative sample, 1 if it is a positive sample, and the image is classified as a fuzzy sample and not included in the training dataset if it falls between these values.

Next, we use the ResNet50 model pre-trained on the ImageNet dataset as the transfer model⁹⁰. The pre-trained model performs well in high-dimensional image feature extraction, allowing the target model to converge earlier and exhibit better generalization performance. During training, the ratio of images in the training set to the test set is 8:2. We first freeze all parameters and train only the fully connected (FC) layer, then unfreeze and fine-tune all parameters. The cross-entropy loss function (Formula 8) and accuracy (Formula 9) are used as performance evaluation metrics to measure the difference between predicted and actual values.

$$\:\begin{array}{c}{L}_{c}=-{\sum\:}_{i=1}^{n}{\mathcal{Y}}_{i}\text{log}\left({\widehat{\mathcal{Y}}}_{i}\right)\end{array}$$

(8)

$$\:\begin{array}{c}\text{A}\text{c}\text{c}=\frac{{N}_{{\widehat{\mathcal{Y}}}_{i}={\mathcal{Y}}_{i}}}{N}\end{array}$$

(9)

In the above equations, $\:n$ represents the total number of classes, $\:N$ represents the total number of samples, $\:{\mathcal{Y}}_{i}$ represents the true label, $\:{\widehat{\mathcal{Y}}}_{i}$ ​represents the predicted label, and, $\:{N}_{{\widehat{\mathcal{Y}}}_{i}={\mathcal{Y}}_{i}}$ represents the number of correct predictions.

Finally, streetscape images from the study area are used as input for the six perception models to obtain scores across different dimensions. The probability of being predicted as a positive category (ranging from 0 to 1) is taken as the perception score, with a higher probability indicating a stronger perception in that dimension.

Fig. 5

Proposed perceived score prediction framework.

Comprehensive index calculation

After obtaining measurement results from both subjective and objective dimensions, we use a normalization weighting formula to calculate the CI. This study introduces the concept of a walkability needs system, where the MAI represents basic needs like accessibility or connectivity at the lower end of the pyramid, and the MII and SPI represent higher-level needs such as comfort (Fig. 6)¹⁴. The CI for a specific point is calculated using the following formula:

$$\:\begin{array}{c}{S}_{ci}={S}_{mai}\times\:\left(\frac{{S}_{mii}}{{S}_{mai}+{S}_{mii}}\right)+{S}_{mii}\times\:\left(\frac{{S}_{mai}}{{S}_{mai}+{S}_{mii}}\right)+{S}_{spi}\times\:\left(\frac{{S}_{mai}}{{S}_{mai}+{S}_{mii}}\right)\end{array}$$

(10)

In the formula, $\:{S}_{c}$, $\:{S}_{w}$, $\:{S}_{m}$, and $\:{S}_{p}$ represent the comprehensive index, Macro-Scale Index, Micro-Scale Index, and Street Walking Preferences Index, respectively. This formula includes two scenarios:

• 1. When the MAI is greater than the MII, it indicates good walkability, and residents’ basic needs are met. In this case, street-level characteristics are of greater interest, so the MII and SPI are given higher weights.

• 2. Conversely, when an area’s service facilities or street network cannot meet residents’ daily needs, attention to street-level landscape elements diminishes, and the Walkability Index is given a higher weight.

This compensatory method allows for a more comprehensive consideration of pedestrians’ actual walking needs and a more reasonable assessment of street walkability^12,43.

Fig. 6

Diagram illustrating the walkability needs system.

Study area and data collection

Beijing is the political, cultural, and educational center of China, and a global metropolis known for its urban scale and economic size. The core feature of Beijing’s urban planning is the “ring road development model,” involving the construction of ring-shaped transportation networks to guide outward expansion. From the 1950s to the early 21st century, in response to traffic pressures and to enhance the economic vitality of new urban areas, Beijing gradually constructed five ring-shaped transportation networks around the city center, transforming the city’s layout from concentration to dispersion and from a single-center to a multi-center configuration⁹¹. This ring road system, similar to those in Moscow, Tokyo, and London, promotes multi-center development and improves traffic efficiency. This study selects the area within Beijing’s Fifth Ring Road as the research area, allowing for an in-depth understanding of the impact of ring road planning on the urban walking environment (Fig. 7), and providing comparable cases and experiences for other similar cities facing urbanization challenges.

We collected four types of basic data within the research area: road network vector data, streetscape image data, POI (Points of Interest) data, and survey data. The road network data was obtained from OpenStreetMap (OSM)⁹². To ensure accuracy, we excluded wide urban expressways with few pedestrians, small paths without streetscape images, and some service-oriented internal roads, resulting in 39,294 road segments after breaking the roads at intersections. Streetscape images were acquired through the Baidu Maps API⁹³. To replicate the natural human field of view (120° horizontal, 60° vertical), we adjusted the API and URL parameters, setting the tilt angle to 6° and the image resolution to 500 × 375. Collection points were set at 50 m intervals along each urban street, capturing streetscape images from four directions (0°, 90, 180, and 270°) at the most recent dates, yielding 78,566 sampling points and 314,264 images. POI data was obtained through the Amap Open Platform API⁹⁴, collecting 220,748 facility points related to residents’ lives within and around Beijing’s Fifth Ring Road. Surveys were collected both online and offline. Online surveys were created using the Wenjuanxing platform and distributed via social media⁹⁵, while offline surveys invited pedestrians in various shopping malls and residential areas within the study area to fill them out on mobile devices.

Fig. 7

Study area within Beijing’s fifth ring road.

Results and analysis

Objective measurement results

Macro-scale index

This study normalized the MAI within Beijing’s Fifth Ring Road and calculated the weights of five types of urban facilities (Fig. 8). The Walkability Index exhibits a significant similarity in spatial distribution to various urban facilities and is largely consistent with the Walk Score reported by Long et al.⁹⁶. The northeastern part of the Fifth Ring Road has a significantly higher MAI than the southwestern part. Except for “Educational and Cultural” facilities, the distribution of other facilities shows a decreasing trend from the Second Ring Road outward.

In the spatial distribution of the MAI, Dongcheng District and Chaoyang District have the highest values. Dongcheng District, as Beijing’s historical center, is rich in cultural facilities. Chaoyang District, being the diplomatic and most populous district, has a well-developed range of living support services and numerous commercial centers, resulting in a very high MAI.

Figure 8b,d highlight similar distribution patterns. Haidian District, despite being in the northwest and distant from the central city, hosts many well-known universities, giving it the highest weight in the “Educational and Cultural” category (Fig. 8b). The area within the Second Ring Road, Beijing’s old core, has many museums and historical sites, driving the development of transportation facilities and leading to higher weights in the “Municipal Services” category compared to other areas (Fig. 8d).

Fig. 8

The normalized the macro-scale index and weighted calculation results of five types of urban facilities.

Logit model fitting results

We conducted a preliminary survey in May 2023 and a formal survey from June to September 2023. The questionnaire consisted of two parts: personal information (gender and age, see Table 5) and scenario selection tasks (twelve choice sets). We recorded the completion time for each respondent and excluded samples completed in less than 30 s. A total of 659 valid questionnaires were collected, exceeding the estimated minimum sample size of 506. The gender and age distribution of the sample is similar to data from the seventh population census of Beijing⁹⁷.

Table 5 Gender and age statistics of sample data and study area.

The questionnaire data were used for model fitting, and the results are shown in (Table 6). The overall model effect is significant, with a likelihood ratio chi-squared (LR chi2) value of 1827.77 and a p-value less than 0.001. The Pseudo R2 is 0.1448, higher than the minimum acceptable standard for a conditional Logit model⁹⁸. Except for the low level of street signs, all attribute levels have p-values less than 0.05. Narrow sidewalks have the greatest impact on walking preferences, while low levels of street signs have the least impact.

Further analysis reveals that less greenery and narrower sidewalks negatively impact pedestrian preferences, while sidewalks with barriers and lower pedestrian flow positively influence preferences. These findings are consistent with empirical knowledge and existing research, as barriers, street trees, and sidewalk width are key elements affecting pedestrians’ sense of security^99,100,101, and lower pedestrian flow improves street occupancy and maintains appropriate social distance¹⁰². Our study also shows that lower traffic flow negatively affects pedestrian preferences, possibly because fewer vehicles, while enhancing traffic safety perception, may reduce crime safety perception. There is currently no consensus on traffic flow’s impact on pedestrian perception. Some scholars associate high traffic flow with noise and safety hazards, reducing pedestrian comfort and safety^103,104, while other studies show a weak correlation between motor vehicle flow and pedestrian walking¹⁰⁵. Further research is needed to clarify this contradiction.

Additionally, we found that open street interfaces positively influence walking perception by reducing feelings of enclosure and oppression^25,57. However, semi-open enclosures can disrupt street continuity and have a suppressive effect¹⁰⁶. The influence of street signs on pedestrians is relatively small, with medium levels of street signs having a more significant promoting effect than other levels^107,108.

Table 6 Questionnaire data for model fitting.

Micro-scale index

Based on the estimated coefficients from the Logit regression model, we incorporated the micro-element data of each point to calculate the MII for the streets. The results are displayed within hexagonal walking units with a radius of 480 m (the product of the walking comfort range and standard walking speed)⁶⁶. Figure 9 shows that the MII within the Fifth Ring Road exhibits a ring-like distribution: lower scores in the city center and higher scores in peripheral areas. Additionally, main roads generally have better evaluations than adjacent local roads. In the old city area, especially within the Second Ring Road, large areas exhibit low walkability. In contrast, areas with high walkability are mainly concentrated near parks and between the Fourth and Fifth Ring Roads.

To verify the validity of the calculation results, we selected two groups of streetscape images based on the extreme distribution of walkability (Fig. 10). The streetscapes in these groups show significant differences. Streets with better evaluations share features such as spacious sidewalks, open spaces, and abundant greenery. In contrast, streets with poorer evaluations generally lack sidewalks and are lined with continuous building facades. The area between the Fourth and Fifth Ring Roads is a newly developed urban extension area with high block connectivity and rich streetscape elements, performing excellently in the evaluation¹⁰⁹. Additionally, streets around most parks and attractions exhibit similar characteristics. In comparison, the area within the Second Ring Road of Beijing is a restricted development zone, where large-scale construction is not feasible, and only partial repairs are possible. Streets within these areas are typically hutongs and alleys, which usually have less greenery and enclosed spaces, leading to poorer evaluation results^91,110.

Fig. 9

Micro-scale index within the Beijing fifth ring road.

Fig. 10

Streetscape comparison between the two extreme walkability groups.

Differences between macro-scale index and micro-scale index

To observe the differences between indicators at different levels and account for spatial distribution heterogeneity, we conducted a bivariate spatial autocorrelation analysis between the MII and the MAI. The bivariate global Moran’s I index is -0.208, with a p-value less than 0.001 and a Z-value of -52.33, indicating significant negative spatial correlation and segregation of walkability scores at different scales. The results of the bivariate local autocorrelation are shown in (Fig. 11). At the 95% significance level, the core urban area within the Second Ring Road primarily exhibits high-low clusters, while low-high clusters appear in peripheral urban areas. This spatial clustering pattern reflects the mismatch between street walking conditions and living service facilities within communities¹¹¹. We will use a normalization weighting method in the CI calculation to mitigate these disparities.

Fig. 11

The bivariate local autocorrelation results.

Spatial distribution of street walking preferences index

The single-dimension score for each point is determined by the arithmetic average of the scores from images in four directions. The final score is the sum of the six-dimension scores (negative perceptions are calculated by subtracting the dimension score from 100). Figure 12 presents descriptive statistics of perception scores across the six dimensions, indicating that all scores follow a normal distribution. Figure 13 visually displays the spatial mapping results of the six subjective perception dimensions.

In the four positive perception dimensions, the distribution patterns of “wealthy,” “lively,” and “safe” show a trend that rises and then falls from the inner city to the outer city, with higher scores mainly between the Second and Fourth Ring Roads. High-scoring areas include Fengtai Science Park in the southwest, Wangjing Central Park and the China International Exhibition Center in the southeast, and the Capital Library in the east, overlapping with high-scoring areas of the MAI. These areas are commercial and economic centers of old urban or near suburban Beijing, characterized by high land-use mix, well-developed infrastructure, and good walking conditions.

The clustering pattern of the “beautiful” dimension shows lower scores in the city center and higher scores in peripheral areas, particularly between the Fourth and Fifth Ring Roads, with the highest scores in the western Laoshan Urban Leisure Park. Moreover,, most major roads have higher overall scores than roads within clusters, similar to the MII distribution. This finding also corroborates the conclusion of Wang et al.¹¹² and Lei et al.¹¹³, who identified an overlap between highly accessible streets (i.e., major roads) and streets with high perceived scores in their studies.

In the two negative perception dimensions, “depressing” scores higher within the Third Ring Road, gradually decreasing outward. High-scoring areas almost completely overlap with high-scoring areas of positive dimensions such as “safe” and “wealthy.” In contrast, “boring” exhibits an entirely opposite spatial distribution pattern.

Based on the above analysis, we obtain preliminary comparative results similar to existing research:

• 1. The area between the Second and Fourth Ring Roads brings more positive perceptions such as “safe” and “wealthy” but also generates more negative cognitions. Pearson correlation analysis (Fig. 14) shows significant positive correlations among “wealthy,” “lively,” and “safe, ” while “beautiful” is negatively correlated with “depressing, ” and “lively” is negatively correlated with “boring, ” confirming this point^114,115.

• 2. There is an intrinsic connection between subjective and objective measurement results in their spatial distribution. High-scoring areas of “wealthy,” “lively,” and “safe” align with high-scoring areas of the MAI, while “beautiful” and “depressing” are similar to the MII.

These conclusions indicate that both macro-level urban form elements and micro-level street factors affect street walkability. The first conclusion is consistent with public health research showing that rapid urbanization has improved living standards but can lead to a sense of oppression from prolonged exposure to a concrete environment^116,117. To better understand the second point, we will use regression models for quantitative analysis in the next section.

Fig. 12

Descriptive statistics of the perception scores across the six dimensions.

Fig. 13

Spatial mapping results of the six subjective perception dimensions.

Fig. 14

Pearson correlation analysis indicates correlations between different perceptions.

Differences and similarities between subjective and objective measurements

In this section, we use multivariable regression analysis to explore the correlation between street elements and perception scores, further analyzing the differences and similarities between subjective and objective measurements. All element data were obtained through the semantic segmentation model. Before regression analysis, we performed multicollinearity checks and excluded variables with a Variance Inflation Factor (VIF) greater than 5. The final regression analysis results are shown in Fig. 15, focusing on variables with a significance level below 0.05 and the top ten absolute values of Beta coefficients. Blue represents a positive correlation, gray represents a negative correlation, and the length of the bar indicates the size of the Beta coefficient.

Figure 15 shows that the contributing factors for each perception dimension vary, but trees and the sky are important predictors across all dimensions. In the “beautiful” dimension, trees and the sky have a strong positive influence, while buildings have a negative impact. This explains why the spatial distribution of “beautiful” is similar to the MII. The streets between the Fourth and Fifth Ring Roads are wider and have greater building setbacks, increasing the proportion of sky and greenery in pedestrians’ view and reducing the proportion of buildings. This results in higher scores for the city outskirts compared to the city center. Additionally, the positive contributing factors in the “beautiful” dimension (trees, cars, and fences) have high weights in the MII, leading to similar distribution patterns.

In the “wealthy,” “lively,” and “safe” dimensions, the same positive factors (cars, roads) and negative factors (sky, buildings) are present. These elements are typical in high MAI areas such as commercial zones, residential areas, and schools. In these areas, the number of vehicles and the proportion of buildings increase, while the sky view proportion decreases. This explains the consistency between the distribution patterns of these three dimensions and the MAI.

Additionally, we find that landscape elements such as the sky, roads, trees, and fences reduce the feeling of “depression” but may also increase the feeling of “boredom.” Previous studies have shown that natural landscape elements enhance psychological health^118,119 and reduce the impact of mental illnesses¹¹⁶. However, the monotony and repetitiveness of these elements may induce boredom due to a lack of sufficient stimuli^120,121. The sky ranks high among the negative factors in the “depression,” “lively,” “safe,” and “wealthy” dimensions, and trees rank high in the first three dimensions. This partially explains the similarity in the perception distribution of these four dimensions.

Fig. 15

Multivariable regression analysis results.

Comprehensive index evaluation

To better present the evaluation results and visually demonstrate the relationship between different land use types and the quality of the walking environment, while also identifying areas of walkability inequality using economic, population, housing, and other data, we mapped the calculated CI results onto a land use classification map (Fig. 16)¹²². The figure shows that Beijing’s streets form a multi-core cluster structure. Financial and cultural centers such as Workers’ Stadium, Wangjing, and Zhongguancun serve as main cores, with secondary centers like Fengtai Wanda Plaza, the National Theatre of China, and Wanfangting Park spreading to surrounding areas. Low CI areas are mainly located on the city outskirts, where commercial and service facilities are underdeveloped, whereas high CI areas are concentrated in the city center. This distribution pattern reflects the imbalance in Beijing’s urban development, with significant differences in land use efficiency between central urban areas and peripheral regions.

A detailed analysis shows that high-scoring areas are supported by commercial plots, supplemented by residential and public plots, forming regions with higher comprehensive indices. Areas with a CI greater than 0.55 (normalized value) are invariably surrounded by commercial land. This indicates that the mix of land use, especially the presence of commercial land, is highly correlated with walkability scores^123,124. Related studies also confirm that public and commercial services are positively correlated with residential land prices¹²⁵.

Further analyzing the descriptive statistics of the CI and land use types (Fig. 17), we find that commercial plots have the highest mean and median, followed by residential and public land, while industrial land has the lowest index. Commercial land, with its concentrated commercial activities and service facilities, typically provides a good walking environment, reflected in the high CI. Residential and public land show significant fluctuations in peak and trough values due to varying types of housing and facilities. Public land includes diverse facilities such as schools, hospitals, and parks, leading to significant variation in the CI. Industrial land, mostly located on the city outskirts, has significantly lower environmental aesthetics and fewer service facilities, resulting in the lowest CI scores.

Fig. 16

The 5th ring road land use classification and composite index maps.

Fig. 17

Descriptive statistics of the composite index on land uses.

Discussion

With rapid urbanization, streetscape has become a primary focus in urban planning. It affects residents’ daily quality of life and represents sustainable development and human-centered urban design. Previous studies on walkability evaluation systems have faced limitations such as low precision, unreasonable weight determination, and limited coverage, resulting in partial evaluations that fail to reflect diverse pedestrian needs. This study integrates previous evaluation systems, leveraging high-precision urban big data and advanced computer vision algorithms, to propose a novel framework for a comprehensive subjective and objective scoring system. This framework overcomes existing evaluation system limitations while reducing resource consumption and deployment costs.

In the context of walkability priority, our evaluation framework has significant implications for street construction in medium to large cities. Comprehensive evaluation results provide crucial spatial information for decision-makers, facilitating precise street management and targeted strategic adjustments to improve urban functions. Additionally, the framework supports cross-disciplinary research in transportation and public health. For example, in transportation planning, it helps identify pedestrian-friendly areas, guiding traffic flow management and optimizing public transportation layouts^40,126. In public health research, the built environment is increasingly acknowledged as a critical factor influencing residents’ health outcomes and lifestyle choices. Assessing street environments not only offers valuable insights into accessibility and mobility but also plays a pivotal role in fostering healthier lifestyles by promoting physical activity and mitigating the incidence of chronic conditions, such as obesity and cardiovascular diseases¹²⁷.

Our study yields several key findings. First, the empirical results demonstrate the reliability of the proposed framework in accurately measuring walkability distribution and revealing social stratification in urban development. The CI results indicate that the walkability environment in the northeastern part of the Fifth Ring Road in Beijing is significantly better than in the southwestern part. Uneven walkability environments suggest potential economic policy biases and inequalities in social benefits such as housing prices, healthcare, and education¹²⁸. Indeed, the northeastern part includes Beijing’s economically most developed Haidian and Chaoyang districts¹²⁹, with significantly higher numbers of permanent residents¹³⁰, schools¹³¹, and top-tier hospitals¹³² compared to the southwestern region, resulting in higher regional vitality levels¹³³. Furthermore, the real estate market activity in the northeastern region continues to rise due to superior social resources¹³⁴.

Second, our empirical research reinforces the theoretical foundation of the evaluation framework. Detailed comparison results show significant spatial distribution differences between the MAI and the MII, with similar distribution patterns observed between the six different perception dimensions and either the MAI or the MII. This indicates that constructing a walkability measurement system requires comprehensively considering interactions between subjective and objective dimensions, as well as the actual needs and preferences of pedestrians in different regions. In urban core areas, there is a high MAI but a low MII, whereas the opposite pattern is observed in peripheral areas. This finding is consistent with Ki et al.‘s study on walkability inequity in the Los Angeles region¹¹¹. Additionally, the MAI correlates with positive pedestrian perceptions of “wealthy,” “lively,” and “safe,” while the MII correlates with “beautiful” and “depressing.” This result is consistent with other scholars’ conclusions, indicating that the effect of objective accessibility on perceived accessibility is minimal^8,135, and higher levels of objective walkability do not necessarily correspond to more positive perceptions of walkability¹³⁶.

Third, from the perspective of walkability assessment, the MAI based on macro elements more accurately reflects the walkability of high-density cities characterized by ring-road traffic development patterns. In the empirical study, we used normalized weights to mitigate differences between different level indicators. However, we observed that the spatial clustering pattern of the CI is significantly similar to the MAI’s spatial distribution. Additionally, the final perception scores align with the MAI distribution pattern. Spatial autocorrelation results show that while there are differences between the MAI and the MII in the city center, the differences are not significant and are less pronounced than in the outer city. Streets with a high MAI typically have medium levels of MII and perception scores, whereas the opposite is unlikely to be true. This finding is consistent with the studies conducted by Nyunt et al. and Kim et al. in Singapore and Seoul, respectively, both of which feature similar ring-road development patterns and high building densities^137,138. However, it contradicts the findings of Koo et al.⁶, likely due to differences in the planning backgrounds of Beijing and Atlanta. The area within Beijing’s Fifth Ring Road follows traditional Chinese planning, characterized by a dense urban structure and mixed land use that promote community-level interactions. In contrast, Atlanta’s urban layout is more dispersed and tends to employ zoning, giving street-level factors a greater influence on human activities.

This study has limitations but also points to directions for future research. First, the calculation of the MAI may be subject to biases in allocating facility weights and selecting correction data. To enhance objectivity, weighting factors should be derived from field investigations. Additionally, spatial syntax parameters, such as integration and connectivity, can be used as correction data. Second, the stated preference survey limited the number of variables included in the calculation of micro-indicators. We only considered seven variables, although more potential elements (such as façade form and public space) significantly affect pedestrian perception¹³⁹. Additionally, we could not include dynamic attributes (such as lighting, honking sounds, and temperature). Future research can employ virtual reality technologies to create more realistic simulation environments¹⁴⁰ or use physiological measurement tools like skin conductance and Electroencephalography. Third, due to dataset limitations, we only selected six labels to represent human perception scores. As more emotional datasets emerge, we can include more perception indicators (such as excitement and comfort). Fourth, the street view images we used have inconsistencies in years and seasons and differ in collection angles from pedestrian perspectives, potentially causing deviations in element recognition and emotional prediction results. Standardizing image collection processes and optimizing algorithms can reduce these biases^140,141. Fifth, the comprehensive walkability evaluation framework proposed in this paper primarily considers individuals without mobility impairments, thereby overlooking populations with disabilities. Furthermore, due to high implementation costs and limitations in data availability, this framework is not suitable for regions lacking comprehensive urban datasets. In future research, we will develop a walking environment assessment framework tailored to special populations—such as people with disabilities, the elderly, or children. We will also enhance the framework’s generalizability to facilitate its application across different urban settings.

Conclusion

This study integrates geographic big data and image recognition technology to develop a data-driven walkability assessment framework, providing a comprehensive evaluation from both subjective and objective dimensions. The proposed framework is cost-effective and highly predictive, making it suitable for large-scale urban walkability assessments. Empirical results reveal spatial imbalances in the CI and its potential correlation with various socio-economic factors, demonstrating the framework’s effectiveness. Further analysis highlights significant differences between the community-based MAI and the street element-based MII in urban core areas versus suburban areas. Objective and perceived walkability are intricately linked, both contributing to street walkability suitability, enriching our theoretical framework. Notably, in high-density urban environments, the MAI more accurately reflects the walkability of specific areas compared to the MII.

Amid rapid urbanization, this study provides new insights for urban planning and public health research, laying the theoretical foundation for promoting sustainable urban development, improving residents’ quality of life, and transforming economic growth models. Future research will incorporate more diverse data (such as traffic flow and video surveillance data) and employ more scientific methods to further refine the current framework.