Based on defining the measurement of street space quality, the measurement systems have mainly revolved around two aspects: physical environment and pedestrian perception. These are generally addressed in Table 1.

In terms of measuring the physical environment of streets, the measurement systems involve one or more of the five categories of street functionality, diversity, serviceability, demography, and architecture (37–39), which have been widely confirmed in academic discussions. However, the applicability of different measurement systems in specific empirical studies remains to be discussed. To effectively assess the impact of physical environmental quality on residents' lives, measurement factors must be selected based on the area's actual situation.

In terms of measuring the pedestrian perception of streets, academic discussions and applications have applied the comprehensive multisensory measurement from sensory stimulation and environment cognition, and the dimensions of measuring perspectives are divided into five categories: psychology, vision, auditory, touch, and olfaction (28, 33–36, 40). However, pedestrian activity behavior may change in a controlled environment, and the results may not accurately represent the sensory feedback in the actual streets (41, 42).

The measurement technologies of physical environment are based on spatial data that include a traditional three-dimension spatial data, location-based services data (LBS), point of interest (POI) data, and transportation data. Yang et al. (43) used LBS and POI data to calculate the hill number in ecology for determining the degree of spatial mixing. Zhang et al. (44) used a multi-source urban geospatial big data to analyze Baidu's POI data and real-time Tencent user density for the degree of crowding. But, these measurement technologies still face some problems, such as difficult data acquisition, high cost, insufficient accuracy, operability of measurement technology, and so on.

The pedestrian perception measurement technologies are classified into three categories: traditional questionnaire survey, emotional analysis of social media data, and physiologic perceptron analysis. Rollero and Piccoli (45) used traditional questionnaires to investigate the environmental perceptions of citizens. Alexander et al. (46) used a random sample of Swedish adults to test 17 issues that were related to local street aesthetics, safety, and society. The following paragraphs discuss the last measurement technologies, respectively.

First, in terms of social media data analysis technologies, some researchers use semantic analysis to analyze the emotional characteristics of pedestrians in social media data such as text classification, sentiment analysis, and intention recognition. For example, Kovacs-Györi et al. (47) used Twitter data to extract the emotions of pedestrians, and analyze the characteristics of emotions over time. Plunz et al. (48) used a Twitter database with geolocation information to analyze the emotional impact of green spaces on pedestrians.

Second, physiological perception analysis is widely using street view images to measure the pedestrian visual perception, and uses a fully convolutional neural network to segment the street view images (49). Several studies utilized biosensors to evaluate the pedestrian response in different street spaces, including Electroencephalogram (EEG), Eye-tracking Metric (ETM), and pressure tests. These technologies provide new technical references for measuring street space quality. For example, Kim et al. (50) used EEG to assess the relationship between the walking environment and pedestrian perception at night.

Therefore, the measurement technologies of street space quality are based on qualitative analysis and quantitative expression. Machine learning technologies are used to analyze the correlation between street space quality and natural environment, social economy, pedestrian physiological characteristics, and other factors. However, whether the current pedestrian perception measurement methods are equivalent to the actual pedestrian perception of walking is still worth discussing.

Academic studies mainly adopt the Experts Grading Method to verify the measurement results. Although the rating scale and classification method have subjective deviations, the result is generally considered acceptable, and is highly effective in certain situations (51). For example, Rezvanipour et al. (36) invited experts in different fields to evaluate the relationship between the physical environment and social attributes of the street from the aspects of urban policy, social behavior, and urban aesthetics. Ewing and Handy et al. (33) established an expert evaluation team to qualitatively define the quality of street space and to score different street scenes. While the results of expert evaluation can represent the standardization of space design, they may not accurately reflect the perspectives of participants in urban sociology and environmental behavior research.

Subjective and objective comparative verification methods are mainly divided into two types: public evaluation and machine learning comparative verification, and expert evaluation and machine learning comparative verification. Both methods use deep learning algorithms such as convolutional neural networks to train sample data sets, and to analyze the consistency between the classification results of the samples and the subjective evaluation results. For example, Liu et al. (52) used the method of expert rating to label images and train the models, then, conducted a field survey on pedestrians and compared the public's rating scores with the machine rating scores. Ye et al. (53) used SegNet to extract the design elements that affect the quality of street space from street view images, produced a Java-based program to collect expert preferences through pairwise comparison, and evaluated the effectiveness of the model through ANN training.

Based on the above discussion, there is a lack of comprehensive subjective evaluation of street space quality from the two levels of expert scoring and public evaluation. In addition, whether the results obtained by training the model only with street view images can reflect the overall quality of street space needs to be verified. Therefore, this study avoids the defects of current effectiveness testing, and comprehensively verifies the effectiveness of the measurement results through six aspects: public evaluation, street classification, expert scoring, data collection, model measurement, and result comparison.

According to the previous studies, this research constructs a multi-dimensional systematic quality measurement model of street space, and comprehensively verifies the effectiveness of the measurement results from subjective and objective.

The old town of Wuchang is located in the central of Wuhan City, with a total area of 7.7 square kilometers (Supplementary Figure 1). It had maintained the first-class administrative center for more than 600 years from the Yuan Dynasty to the Qing Dynasty. The old town has a clear road network (Supplementary Figure 2), but the north-south dual-city pattern is naturally formed due to the obstruction of Snake Mountain. Some traditional street textures are retained in the north, and new urban roads are mainly in the south.

This study applies Baidu Maps API to collect the architecture data, street network data, POI data, and heat map data. The ArcGIS analysis platform processes the original architecture data and simplifies the street networks (Supplementary Figure 3). The 62 streets and 215 street sections were determined as research objects. Concerning the classification of national economic industries, regional functions, and the development of the old town, this study selects 3,066 POI data including five categories: commercial services, public management and public services, transportation facilities, residential buildings, and green spaces. This study collects the 112 rasterized heat maps of each hour from 7:00 to 23:00, between June 1 and 7 in 2021. And, by calculating the average heating value at different time points in each street section, the clustering of active people in the street space is evaluated.

According to the Detailed Regulations for Wuhan Urban Design Management, the neighborhood units in the old town are divided into three levels. The sampling points for collecting street view images are selected according to the different levels of the neighborhood (Supplementary Table 1).

This study uses the DeepLab V3 model (54) trained on the Cityscapes dataset (55) to semantically segment the panorama image for obtaining the proportion of street elements. Limited by the computing device and the size of the collected panorama image dataset, the pre-trained DeepLab model is finetuned on our street view dataset using a transfer learning strategy and achieves mIoU of about 78.8%. The finetuning process is implemented on the open-source deep learning framework TensorFlow (56). Some studies (57, 58) have also illustrated that this segmentation model performed well in Chinese cities. It can segment the panorama into 20 categories, such as sky, road, vegetation, and vehicles, to achieve a pixel-level image segmentation and extract various street space elements (Figure 2). Through the statistical analysis of the segmentation results, the proportion of research elements is obtained, so the segmentation results corresponding to each street scene sampling point are connected with the street section.

When constructing the measurement model, this study mainly expounds from three aspects: index selection, selection reasons, and calculation methods. First, the measurement dimensions include the physical environment and pedestrian perception according to the concept of street space, while the specific measurement levels, indicators, and factors are determined by nature, culture, architecture, economy, infrastructure, and residents. Then, when considering the specific measurement level, a secondary selection of measurement indicators is carried out based on the world's established indicators and literature studies. Finally, the selected measurement indicators are subdivided to define the measurement factors and quantitative methods.

At the dimension of physical environment, many quality measurement systems of street space are aimed at the material space elements. Given the comprehensiveness of the measurement and the comparison of other measurement index systems, the 5D index proposed by Ewing and Cervero has been widely confirmed in the built environment measurement research (29). So, this study focuses on the Design, Density, Diversity, Destination accessibility, and Distance to transit in the 5D index as the physical environment dimension measurement levels in the quality measurement model (24).

From the pedestrian perception dimension, this study clustering literature about street space vitality and pedestrian perception from 2001 to 2021 by CiteSpace software show results that research on the subjective perception of street space mainly focuses on streets' sociality, safety, and status (Supplementary Figure 4).

Based on the above analysis, this study determined the measurement levels in the physical environment dimensions are 5D (Design, Destination accessibility, Distance to transit, Density, and Diversity). Then, the 3S pedestrian perception levels are Sociality, Safety, and Status. In the 5D+3S measurement model, a series of representative measurement index lists are selected to comprehensively evaluate the spatial quality of streets (Table 2).

There are 20 measurement factors in this study, and the properties and dimensions between the factors are different. When the values of the index measurement differ significantly, the impact of the high-value index on the comprehensive analysis increases, resulting in a deviation of the results. Therefore, this study selects the min-max data standardization method to re-scale the raw data without changing the original data structure.

To avoid the inaccuracy of the index assignment results caused by subjective evaluation method factors such as personal emotion and cognitive bias (70, 71), this study uses the entropy weight method to objectively assign the index weight. The entropy weight method (EWM) is a well-known and widely used information weight model. In comparison to other subjective weight models, the primary advantage of EWM is that it eliminates human factor interference with index weights, thereby increasing the objectivity of the results of the comprehensive evaluation (81). The degree of differentiation is used to determine the value in EWM. The greater the dispersion of the measured value, the greater the differentiation of the indicator, and thus, the more information that can be derived.

Then, the entropy weight method is applied to calculate the weight (λ_i) of each index to calculate the Comprehensive Value of Street (CVS). According to the weight of each measurement index, the comprehensive quality λ_i measurement value formula of street space is as follows: CVS=0.254×GVI+0.252×SO+0.104×SI+0.378×SC        +0.055×TA+0.035×CA+0.114×FA +0.095×DS        +0.052×DB+0.033×HW+0.036×HR+0.044×FR        +0.401×NPF+0.486×MPF +0.170×PC+0.073        ×+IS0.004×TS+0.382×HCL+0.238×CA+0.133        ×LA

To verify the effectiveness of the proposed 5D+3S quality measurement model of street space, a hierarchical clustering algorithm is applied to automatically analyze each street index data without human intervention by automatically dividing the street data with similar spatial quality into a cluster. This clustering algorithm is conducted with open-source Python libraries, SciPy, and Scikit-learn. In the initial stage, each record in the street space dataset is regarded as a separate cluster. Then, these small clusters are hierarchically merged into bigger clusters according to the distance criterion with the iteration of the clustering algorithm. In this study, considering the comparison results of some common similarity measurement methods, such as single, average, Ward, and so on, we found that the clustering results obtained by the Ward method are the most content with our data characteristics. In addition, some existing studies (82, 83) also prove that the Ward method had a better performance and robustness. Thus, the Ward minimization variance method is adopted to implement the clustering algorithm, which is formulated as follows: d(u,v)=|v|+|s|Td(u,s)2+|v|+|t|Td(u,t)2-|v|Td(s,t)2 where s and t are clusters, u is a new cluster composed of s and t, and v is an unused cluster, T = |v| + |s| + |t|, and |*| represents the number of data records in the cluster.

According to the concept of hierarchical clustering algorithm, the hierarchical clustering algorithm does not require a predefined number of clusters, and the number of clusters does not affect the final result of the validity test. So, this study clusters the CVS into three categories (good, medium, and poor) as a simple example of 5D+3S measure model validity test. All the data records are assigned with corresponding category labels, taking the three largest clusters as the final clustering results. Then, further analysis of the original results of the proposed 5D+3S measurement model is implemented through other visualization methods. Furthermore, the clustering results and the comprehensive values of the proposed space quality measurement model are also compared to analyze whether the measurement index values in the proposed model are effective.

Using the selected sample streets as the test objects, this study uses an expert scoring and public sentiment polarity in the social media big data to determine the subjective evaluation results. The CVS is taken as the result of objective verification, which is calculated using the 5D+3S quality measurement model of the street space. By comparing the two verification results, to determine whether the measurement model is consistent with subjective evaluations of people and whether it is capable of reflecting the problems with the street space quality from the perspective of the public.

To avoid the cognition bias caused by the randomly selected residents to score the streets, this study collects the Weibo check-in data, public comment data, and Zhihu Q&A data from June 2020 to December 2020 in the central area of Wuhan, and applies a sentiment analysis to evaluate the quality of different street space from public perception. The public evaluation data are processed by two steps: the high-frequency words extraction according to TF-IDF (84) and the sentiment analysis (85, 86) based on the high-frequency words. Then, the three streets with different sentimental evaluations are selected as sample streets for expert scoring, and the scoring criteria are supported by Street Planning and Design Guidelines in Wuhan. In addition, this study collects the data of physical environment and pedestrian perception about sample streets for objective measurement in 5D+3S model. Finally, the result of 5D+3S measure model, public evaluation, and expert scoring are compared to analyze the relationship between these three results. The specific technical framework is shown in Figure 3.

The following results are obtained on the 5D+3S measure of the 5D dimensions in the old town:

Therefore, in the 5D dimensional measurement of the physical environment in the old town streets, the weak spatial service capacity, low traffic efficiency, and imperfect public service facilities system are the main problems.

According to the 5D+3S quality measurement model of street space, the following results are the 3S dimensions of pedestrian perception:

Therefore, the main problems are the insufficient protection of the overall street pattern of the old city and the lack of protection and display of local cultural resources.

In the clustering process, the dendrogram of the hierarchical clustering algorithm is shown in Figure 8A, where the ordinate represents the distance between the linked clusters, and the abscissa represents the number of data points of CVS contained in the cluster represented by each node.

According to the hierarchical clustering result, Class 1 contains 62 data records, Class 2 contains 45 data records, and Class 3 contains 108 data records. In order to illustrate the clustering results more clearly, a t-SNE (90) data visualization technology is applied to reduce the dimension of the original data samples to 2D data. It can be found from Figure 8B that the projection of all the street data records on the two-dimensional plane presents three clear categories, which indicates that the distribution of these samples corresponds to the three clusters in the results of a clustering algorithm.

Then, based on the results of the hierarchical clustering algorithm, the average values of each measurement index value for the street space samples in the three categories are calculated. The radar chart depicting the average values is shown in Supplementary Figure 5A. There are significant differences and separations amongst the index values of the street samples in the three clustered categories. Additionally, Supplementary Figure 5B compares some actual street maps from the three categories of street samples. The results show that the clustering results match the human perception of street space quality.

Furthermore, as illustrated in Figure 9, the CVS distribution is compared to the clustering results. It can be seen that the qualitative outcome of the hierarchical clustering algorithm and the quantitative values of the measurement index obtained using the entropy weight method can also be quite similar. High-quality streets are those that have a high degree of comprehensiveness. Those with a low comprehensive value, on the other hand, are classified as low-quality.

As shown above, the proposed 5D+3S quality measurement model of street space can simultaneously grade the street quality and assign a quantitative quality score to these streets, which can be used by practitioners. The results of the preceding comparative visualization experiments also demonstrate that the proposed 5D+3S measurement model is reasonable and capable of accurately determining the street space quality.

As mentioned in section Subjective and Objective Comparative Validity Test, according to the sentiment analysis and hot word extraction on the text database of public evaluation, all the streets can be divided separately with positive, negative, and neutral sentiments. The top 5 streets in popularity are selected as the studied samples, as listed in Supplementary Table 2. Among them, the street with the highest evaluation is Zhongshan Road, and the street with the lowest evaluation is Luoyu Road. However, the medium-quality streets without good or bad evaluation include Sanyang Road, Nanhu Road, and Youyi Avenue. For the sake of fairness, Youyi Road is selected randomly to represent medium-quality streets. The final three sample streets for expert evaluation are Zhongshan Road, Luoyu Road, and Youyi Road.

First, in terms of expert scoring, this study selected ten local students with a master's degree and ten practitioners in urban planning to establish an expert evaluation team. The specific scoring criteria are based on the following guidelines: Street Planning and Design Guidelines in Wuhan Elements of a Complete Streets Policy, Global Street Design Guide, and the existing highly reliable expert scoring standards (53, 91, 92). According to the scoring criteria shown in Table 3, the quality of the selected sample street is scored from physical environment and pedestrian perception.

Then, this study collects the relevant data about the three streets for objective verification, and the CVS of sample streets are calculated using the 5D+3S measurement model. The results of public evaluation, expert scoring, and CVS are comprehensively compared. It can be seen from Table 4 that the results of public evaluation, expert scoring, and 5D+3S measurement are consistent within a certain error range. The expert scoring of the street with good public evaluation is also high, and the measurement value obtained by the measurement model is also high. This consistency shows that the results of the measurement model are the same as the subjective evaluation of an individual, which proves the effectiveness of the proposed 5D+3S model.

Therefore, the measurement model can reflect the influencing factors of subjective judgment in evaluating the quality of street space. Meanwhile, the measurement model has no human intervention in the calculation, which can exclude the impact of human subjective emotions on the quality judgment and make the measurement results more objective and convincing.

Although this study only selected three streets in the subjective verification, the research points were sampled from these three streets according to a certain distance interval. Therefore, the total number of sampling points in the verification process is relatively sufficient, which can verify the effectiveness of the measurement model to a certain extent.

Therefore, the spatial quality levels corresponding to the comprehensive values of different street space quality measures are determined according to the comparison results, and the specific corresponding relationships are shown in Table 5.