Quantifying the Spatial Ratio of Streets in Bei**g Based on Street-View Images

Abstract

The physical presence of a street, called the “street view”, is a medium through which people perceive the urban form. A street’s spatial ratio is the main feature of the street view, and its measurement and quality are the core issues in the field of urban design. The traditional method of studying urban aspect ratios is manual on-site observation, which is inefficient, incomplete and inaccurate, making it difficult to reveal overall patterns and influencing factors. Street view images (SVI) provide large-scale urban data that, combined with deep learning algorithms, allow for studying street spatial ratios from a broader space-time perspective. This approach can reveal an urban forms’ aesthetics, spatial quality, and evolution process. However, current streetscape research mainly focuses on the creation and maintenance of spatial data infrastructure, street greening, street safety, urban vitality, etc. In this study, quantitative research of the Bei**g street spatial ratio was carried out using street view images, a convolution neural network algorithm, and the classical street spatial ratio theory of urban morphology. Using the DenseNet model, the quantitative measurement of Bei**g’s urban street location, street aspect ratio, and the street symmetry was realized. According to the model identification results, the law of the gradual transition of the street spatial ratio was depicted (from the open and balanced type to the canyon type and from the historical to the modern). Changes in the streets’ spatiotemporal characteristics in the central area of Bei**g were revealed. Based on this, the clustering and distribution phenomena of four street aspect ratio types in Bei**g are discussed and the relationship between the street aspect ratio type and symmetry is summarized, selecting a typical lot for empirical research. The classical theory of street spatial proportion has limitations under the conditions of high-density development in modern cities, and the traditional urban morphology theory, combined with new technical methods such as streetscape images and deep learning algorithms, can provide new ideas for the study of urban space morphology.

1. Introduction

As an important linear public space in the city, a street is the basic skeleton of urban spatial organization [1]. It is also the main place for people to perceive urban form and structure. The physical space form of a street, called the “streetscape” [2], is the core element in the street design. Its scale has an important impact on the urban form. The measurement and quality of the streetscape have recently been central topics in the field of urban design [3]. With the increasing abundance of spatial data and geographic information, geographic information system (GIS) and remote sensing techniques have been widely used in extensive analysis at an urban street level. They can extract streetscape features from the building footprint, street length, tree canopy map**s, and other data, applying them to related research [4,5]. At present, many measurable design features of streetscapes, such as streetscape skeleton variables [6] and scene elements, have been proposed [7]. The main features include street spatial ratio/openness and green rate [8].

“Street space ratio” is an objective description of the relative scale of the street space form, which is related to the buildings on both sides of the street [9]. Street space ratio can be divided into two core indicators: street aspect ratios and street symmetry [10]. Street aspect ratios (Street Width/Street Height, D/H) are the ratios of the street pavement width to the average height of the building interface on both sides of the street from the perspective of the street cross-section. From the perspective of street space aesthetics and humanism, Yoshinobu Ashihara [11] proposed that when street aspect ratios are between 1 and 2, the scale of street space is balanced, and people are provided with a sense of intimacy. When street aspect ratios are 0 to 1 or 2 to 4, the spatial scale of the street is generally interpreted as being too narrow or open. When the street aspect ratios are greater than 4, the street spatial scale is too broad [12]. Furthermore, the street space ratio can be further described by the street symmetry, that is the similarity of the building heights on both sides of the street [13], which can affect the visual perception and thermal comfort of pedestrians [14].

Based on this, we identify and classify three spatial indicators of Bei**g’s urban streets: geographical location, street width-to-height ratio, and street symmetry, using a deep learning algorithm model from a human-centred perspective. Based on the identification results, the spatial proportions and distribution patterns of urban streets are sorted and summarized, analyzing the influencing factors and historical and cultural characteristics. Finally, we select a typical district for empirical research in order to ascertain the perception, design, and control of urban form.

2. Literature Review

In quantitative studies of existing streetscapes, GIS is a relatively classic and mature two-dimensional plane analysis platform. It is usually combined with the street space data composed of vector road network data and building data to form a quantitative and visual analysis method of street space composition elements [3]. Harvey [15] measured 12 characteristics of a street, including length, width, cross-sectional proportions, and street wall continuity using GIS and spatial street data. These characteristics were clustered into four categories of streetscape skeletons: upright, compact, porous, and open [15]. Based on Harvey’s method, the maximum section method of streets was proposed to quantify the continuity of the street interface. The correlation analysis found that the interface continuity was significantly positively correlated with building density, floor area ratio, and road network density, and significantly negatively correlated with road width [16]. GIS data and methods overcame the shortcomings of traditional methods of manual data collection and analysis for small areas or individual streets. However, due to the lack of human perception perspective in these GIS data and methods, the streetscape features extracted from the pedestrian perspective are missing key information.

The emergence of street-view images compensates for this disadvantage. Driven by the proliferation of large-scale image platforms (the coverage and development of services like Google Street View), advances in machine learning and computer vision (capable of automatically extracting a variety of information), and growing computing power (to facilitate the processing of large volumes of images) [3], street view images have gained strong momentum in urban research. On one hand, SVI can objectively and completely reflect the spatial morphological elements of the street. With characteristics like easy access, fast updates, and wide coverage [17], it has rapidly become an important data source for street space research. On the other hand, it is possible to realize automated street spatial elements and index recognition, using image recognition technology represented by Convolutional Neural Network [18].

Street view images provide a valuable source of large-scale urban data, often replacing field visits with virtual audits and the ability to examine visual features from a human (horizontal) perspective [19,20]. This reflects the material properties of the human perspective of the street, which is not possible with traditional methods of studying urban street morphology and other common data sources (aerial or satellite imagery). However, few studies have applied this method to street spatial ratio. The main area of application in current research on SVIs (using image recognition techniques) is the creation and maintenance of spatial data infrastructures [3], that is, to collect spatial data purely for research purposes. In terms of research topics, there are many studies on street greening with green vision rate, greening visibility, and other quantitative indicators, mainly based on the data of street view images. They used multi-spectral remote sensing images, SegNet, and other tools to conduct quantitative research on street greening in large cities such as New York and Shanghai [21,22]. Other studies have extracted street landscape features through street view images, explored street safety [23], classified street functions [24], assessed street quality [25], and discussed urban vitality [26].

SVIs are well-suited for assessing the characteristics of the built environment of streets [27] and are powerful sources for measuring the perception of urban form by pedestrians [28,29]. If the goal is to evaluate a profile of the street aspect ratio and to understand what people see on the ground, street view images offer unparalleled advantages that cannot be achieved by most remote sensing methods [30]. New research has demonstrated the feasibility of measuring street continuity and architectural landscape factors based on deep learning and GSV images [31]. Hu et al. used DenseNet to quantify multidimensional measurements of street aspect ratios, symmetry of buildings on both sides, and complex geographic locations in high-density cities such as Hong Kong [13]. This demonstrated that useful information can be extracted from street view images via deep learning and can support quantitative studies of street spatial proportions. However, current street scale related studies are mainly applied to urban climate analysis, such as the effect of building density on microclimate [10,13], estimating solar radiation and light pollution [32,33,34], air pollution [35,36], and measuring the number of shadows in outdoor recreational spaces [37].

The traditional method of studying urban aspect ratios is manual on-site observation, which is inefficient, incomplete and inaccurate. It is difficult to obtain global conclusions, portray the overall characteristics of urban aspect ratios, and explore the historical patterns of urban morphological evolution and deep-seated influencing factors. As an emerging data source, Street View Imagery can provide a comprehensive, fast and accurate overall characteristics of urban aspect ratios. As an important characteristic parameter of urban morphology, the proportion of urban streets under the SVIs method reflects the overall urban morphological aesthetics and tests the quality of contemporary urban space under the traditional street aesthetics theory. It also reveals the different spatial and temporal characteristics at the level of urban street morphology at different stages of urban historical development and reflects the historical and cultural process of urban morphological evolution. However, there is a lack of relevant studies, with this study aiming to fill this gap.

3. Methods

This study has four main phases, including data acquisition and processing, indicator selection, results of identification, and discussion and implications (Figure 1). Firstly, we obtained vector road network data through Open Street Map (OSM) and used Python to access the Baidu Street Map application interface to obtain street images, and secondly, we realized the quantitative measurement of three indicators: street location, street aspect ratio, and street building interface symmetry, through DenseNet model. The model identification results were aggregated to the corresponding road centerline using Arcgis tools to sort out and summarize the spatial proportional characteristics and distribution patterns of streets in central Bei**g. Finally, in order to explore the influencing factors in depth, four street types are analyzed and a typical lot is selected for empirical evidence.

3.1. Study Area

We selected the central area within the Third Ring Road [38] in the main urban area of Bei**g as our study area. It covers a total area of 159.83 square kilometers. The old city (the area within the Second Ring Road, which is 62.5 square kilometers) consists of the inner city and the outer city on the south side, as well as a vast modern functional area. Through comprehensive research and statistics of Bei**g streets, we found that the central area within the Third Ring Road is rich in street types with a clear time span, which can represent the morphological characteristics of Bei**g streets and reflect the historical and cultural process of urban spatial development. Additionally, the coverage of SVIs is more complete and the collection time is updated.

Bei**g is a city formed by the gradual expansion of the old city in a circle style, and the street space has obvious locational characteristics. To further portray the differences in street morphology for different districts, our study area was divided into five zones: the imperial city, the historic district, inner city built-up area, outer city built-up area, and the second-to-third rings area (Figure 2). We then analyzed the street aspect ratio characteristics in different spatial and temporal contexts. The Imperial City is located in the center of the old city of Bei**g. It is the center of the city, holding the Forbidden City and its affiliated government offices, altar and temple complexes, and royal gardens, covering an area of about 7 square kilometers. It was built during the Yuan Dynasty and developed in the Ming and Qing Dynasties [39,40]. The historical district is composed of 18 historical blocks, including Shichahai, Nanluoguxiang, and Dongjiaoming Alley in the inner city and the outer city, with an area of about 17 square kilometers. It is characterized by narrow alleys formed in the Ming and Qing dynasties and the texture of traditional low-rise quadrangles [38,41]. The inner city built-up area and outer city built-up area refer to the areas of the inner city and the outer city, reflecting the modern urban texture in the old city. The second-to-third rings area is an urban section developed under the guidance of Bei**g’s first master plan (1950–1957) after the founding of the People’s Republic of China [39,42]. This laid the foundation for the modern Bei**g urban development model.

3.2. Data Acquisition and Processing

The vector road network data is urban two-dimensional road data, which is the basic data for the selection of sampling points of street view pictures, the visual expression of street spatial proportion, and the analysis of spatial distribution characteristics [43]. We compared the data update time and the quality of the street view picture and selected Baidu Street View to carry out the spatial proportion research on the streets in the central area of Bei**g Third Ring Road. We then obtained the vector road network data in the central area of Bei**g Third Ring Road through Open Street Map. Due to the excessive details of the original road network, topology errors were easily generated in the subsequent topology analysis. Therefore, the data set of the Bei**g road’s centerline was obtained through road simplification, topology processing, and other processes. It was necessary to input the longitude and latitude coordinates of the sampling points of the street view image to obtain its data. Through literature review, we found that when the distance between the sampling points of the street view image was 50 m, the images were of good quality [44]. However, there were a few cases where the homogeneity of the recognition results was too high, or the street features were omitted. Baidu Street View Map 2021 application interface was called through Python, and four angular parameters were input: latitude and longitude (LAT, LON), horizontal field of view (FOV) 120 degrees, street view camera heading angle (HEADING) (set to 0, 90, 180, 270 degrees), and pitch angle (PITCH) (set to 0 degrees to cover all street graphics under flat view angle). A total of 34,852 sampling points were used, and 139,408 street images were obtained in the central area of Bei**g’s Third Ring Road.

3.3. Indicator Selection

Based on the principles of objectivity, universality, quantifiability, and comparability, the three indicators of street location, street width to height ratio, and street symmetry were selected to quantify the spatial ratio of streets. First, the model was used to classify and identify the geographical location of the streets. They were divided into six categories: general streets, intersections on viaducts, intersections under viaducts, non-intersections under viaducts, and sound barriers. This was done to facilitate subsequent screening of streets with universal research significance for spatial scale analysis. Secondly, the classical street space index “width to height ratio” proposed by Yoshinobu Ashihara was used to identify the street form, and the results were used to classify the streets into four types: 0 < D/H < 1 (canyon-type streets), 1 < D/H < 2 (balanced streets), 2 < D/H < 4 (spacious streets), and D/H > 4 (open streets). Third, the spatial ratio of streets was further described by the street symmetry index and divided into three categories: H1 = H2, H1 > H2, and H1 < H2. The positive left interface of the street is H1, and the right interface is H2. The north direction of the north-south street is the positive direction; the east direction of the east-west street is the positive direction; the northeast direction of the northeast-southwest street is the positive direction; the northwest direction of the southeast-northwest street is the positive direction.

3.4. Convolutional Neural Networks

Convolutional Neural Networks (CNN) is one of the most effective, widely used network models in deep learning and can effectively extract features and output classification results based on different features [45]. In recent years, with the upgrading of hardware and the simultaneous development of software, CNNs have been utilized in many fields, such as image recognition, semantic segmentation, and target detection.

In this paper, a DenseNet deep network model was built to recognize the spatial scale of streets in central Bei**g based on streetscape image data. The method normalized the streetscape image data into a training set and a test set, input the processed training set data into the built DenseNet deep network model for training, and then input the unlabeled data into the model for prediction to obtain the final recognition results.

The street view image data was in RGB format. The images were normalized to a pixel value domain of 0–1. The DenseNet spatial scale recognition model mainly consisted of three sub-models based on location discrimination, aspect ratio discrimination, and symmetry discrimination, which recognize three quantitative metrics in image data (Appendix C). Streetscape images have a large amount of data and distinct features, so supervised learning was selected to train the model. The model structure was divided into an input layer, an implicit layer, and an output layer (Figure 3). The input layer fed the normalized processed street view image data into the model and processed it according to RGB channel dimensions. The implicit layer consisted of Convolutional Layer, Pooling Layer, and Fully-Connected Layer. The Convolutional Layer contained one or more Convolutional Kernel matrices, which were used to extract features from the input image data by convolutional computation and input the Feature Map matrix. The Pooling Layer followed the Convolutional Layer and performed feature selection and redundant information filtering on the feature map. The Fully-Connected Layer arranged all matrices into a column, multiplied each value with the corresponding weight, where the weight value was determined according to the model learning. It then summed them up and connected them to the output layer to obtain the model recognition classification result. Twenty (20) percent of the total number of street images (27,882 street images) were randomly selected for classification and labeling based on street location, street aspect ratio, and street symmetry. The DenseNet model was trained by inputting labeled data, and the model improved the recognition accuracy by analyzing different labeled types of data to achieve recognition of unlabeled data.

4. Results

The accuracy of the model results for the Bei**g streetscape images was counted, the results are shown in

Table 1. Statistical table of model accuracy.

Arrangement	Model Accuracy	Result Type	Accuracy
Street location	74.1%	General street	73.9%
		Intersection	78.7%
		On-viaduct	75.8%
		Crossroads under the viaduct	74.8%
		Non-crossroads under viaducts	71.5%
		Acoustic barriers	76.1%
Street aspect ratio	81.68%	No D/H	91.4%
		0 < D/H < 1	90.2%
		1 < D/H < 2	78.5%
		2 < D/H < 4	75.6%
		D/H > 4	72.7%
Street symmetry	72.4%	H1 = H2	73.2%
		H1 > H2	71.4%
		H1 < H2	72.6%

. The overall accuracy of the model was 76.06%, and the accuracy of all three levels was higher than 70%. The remaining normalized street image data were input into the training model to identify the spatial scale of streets. To ensure the accuracy of the street spatial scale recognition results based on the streetscape images and the DenseNet algorithm model, we screened two types of sampling points: intersection and under-viaduct intersection. We then selected four types of sampling points (general street, on-viaduct, under-viaduct non-intersection, and sound barrier) and used the Spatial Join tool in ArcGIS. The results of the location, street aspect ratio, and symmetry of street building interface were aggregated to the centerline of the road, obtaining the final results.

4.1. Overall Street Aspect Ratio as in Central Bei**g

The total length of streets in the central area of the Third Ring covering the streetscape image collection was 1742.61 km, and different colors were assigned to different types of streets with different aspect ratios in the area for visual representation (Figure 4). The streets in the region showed the general characteristics of predominantly canyon-type streets (0 < D/H < 1), followed by balanced street (1 < D/H < 2), spacious street (2 < D/H < 4), and open streets (D/H > 4), which gradually decrease in length and number. In the central area (from the Imperial City to the Third Ring Road), the proportion of canyon-type streets gradually increased in each district and eventually became dominant. The proportion of balanced streets in each district showed the opposite process of gradual decrease; the proportion of spacious streets and open streets was lower in every district except the Imperial City, the ratio showed a decreasing trend, but the change was not significant (

Table 2. Length and Proportion of Streets with Different Aspect Ratio Types in Central Urban Area and Each District.

Street Aspect Ratio Type	The Imperial City 6.79 km2	The Historic District 15.98 km2	Inner City Built-Up Area 15.59 km2	Outer City Built-Up Area 21.67 km2	The Second to Third Rings Area 96.58 km2	Central City 159.83 km2	Typical Pictures
Canyon-type streets 0 < D/H < 1	8.06 km	92.91 km	107.48 km	134.5 km	668.57 km	1011.5 km
	21.70%	38.41%	56.46%	62.92%	63.10%	58.05%
Balanced streets 1 < D/H < 2	18.18 km	115.82 km	55.41 km	52.75 km	202.36 km	444.52 km
	48.94%	47.89%	29.11%	24.68%	19.10%	25.51%
Spacious streets 2 < D/H < 4	4.35 km	17.66 km	13.93 km	11.99 km	59.61 km	107.54 km
	11.71%	7.30%	7.32%	5.61%	5.63%	6.17%
Open streets D/H > 4	6.02 km	10.83 km	7.23 km	5.77 km	29.41 km	59.26 km
	16.20%	4.48%	3.80%	2.70%	2.78%	3.40%
No aspect ratio street No D/H	0.54 km	4.64 km	6.32 km	8.75 km	99.52 km	119.77 km
	1.45%	1.92%	3.32%	4.09%	9.39%	6.87%
Total	37.15 km	241.86 km	190.37 km	213.76 km	1059.47 km	1742.61 km
	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%

).

4.2. Spatial Characteristics of Streets with Different Aspect Ratio Types

Combined with the recognition results of the convolutional neural network model, the spatial characteristics of different types of streets with different width-to-height ratios in the Third Ring Road of Bei**g were systematically carved from three features: spatial distribution, orientation, and symmetry.

4.2.1. Canyon-Type Streets (0 < D/H < 1)

Canyon-type streets were the most dominant street type within the central city of Bei**g, with a total length of 1011.52 km, accounting for 58.05% of the total length of streets within the central city. As the district moves from inside to outside, the length of canyon-type streets gradually increased as a percentage of the total length of streets in each district, and the growth rate gradually slowed (Appendix B).

The canyon-type streets are spatially polycentric (Figure 5), and the spatial distribution map showed that they are mainly concentrated in the inner city, outer city built-up areas, and the second and third ring roads. Its gathering form was divided into two categories. One was irregular ring-shaped strip form, larger area, mostly gathered around high-grade roads and business centers, such as the East and West Second Ring Road, Xuanwumen West Street, Chaoyangmenwai Street and Finance Street, Jianguomenwai CBD area, etc. (Appendix A). The other was an elliptical scattered distribution, mostly in commercial areas, high-rise residential areas, and university campuses. The hutongs in individual historic districts, such as the Liulichang district, were too narrow and also showed canyon patterns.

Corresponding to the square pattern of Bei**g city, canyon-type streets in the central city were mainly oriented east-west and south-north. The two types of streets accounted for 92.6% of the total length of canyon-type streets, with a slightly higher percentage of east-west oriented streets. The southeast-northwest and northeast-southwest streets were scattered, mostly located near the urban water system, with a slightly increasing proportion from the inside out, forming clusters in the northeast corner of the second-to-third rings and the northwest corner of the second ring (

Table 3. Canyon-type streets orientation and symmetry statistics.

	Street Orientation	The Imperial City	The Historic District	Inner City Built-Up Area	Outer City Built-Up Area	The Second-To-Third Rings Area	All Regions
Street orientation	South-North	3.89 km	43.47 km	59.88 km	60.32 km	319 km	486.56 km
		47.21%	41.05%	48.86%	41.76%	46.97%	45.89%
	East-West	4.24 km	57.82 km	57.45 km	78.09 km	297.68 km	495.28 km
		51.46%	54.60%	46.88%	54.06%	43.83%	46.71%
	Southeast-Northwest	0.03 km	2.43 km	2.6 km	3.48 km	33.96 km	42.5 km
		0.36%	2.29%	2.12%	2.41%	5.00%	4.01%
	Northeast-Southwest	0.08 km	2.17 km	2.63 km	2.57 km	28.48 km	35.93 km
		0.97%	2.05%	2.15%	1.78%	4.19%	3.39%
Street Symmetry	H1 = H2	4.16 km	42.60 km	60.20 km	61.42 km	268.96 km	437.34 km
		51.61%	45.85%	56.01%	45.67%	40.23%	43.24%
	H1 > H2	2.11 km	24.35 km	22.20 km	36.08 km	199.99 km	284.73 km
		26.18%	26.21%	20.66%	26.83%	29.91%	28.15%
	H1 < H2	1.79 km	25.96 km	25.08 km	36.99 km	199.62 km	289.44 km
		22.21%	27.94%	23.33%	27.50%	29.86%	28.61%

).

In terms of street symmetry (Table 3), the length of the “H1 = H2” type of streets accounted for 43.24%. With the location from inside to outside, the proportion of “H1 = H2” type street length gradually decreased and the proportion of “H1 < H2” and “H1 > H2” type gradually increased. Overall, the three types of streets showed spatially uniform distribution characteristics.

4.2.2. Balanced Streets (1 < D/H < 2)

The total length of balanced streets was 444.52 km, with a length ratio of 25.51%. The spatial distribution map showed that the balanced streets were distributed in the old city. They were most concentrated in the northern part of the inner city, along Fuchengmennei, ** to perceive, design, and control urban spatial forms more scientifically and effectively.