Identification and Evaluation of Representative Places in Cities Using Multisource Data: Focusing on Human Perception

Abstract

Discovering the Representative places (RPs) of a city will benefit the understanding of local culture and help to improve life experiences. Previous studies have been limited in regard to the large-scale spatial identification of RPs due to the vagueness of boundaries and the lack of appropriate data sources and efficient tools. Furthermore, human perception of these places remains unclear. To address this gap, this research adopts a novel approach to identify and evaluate the RPs of a city from the perspective of human perception. Our methodology involves the utilization of deep learning systems, text semantic analysis, and other techniques to integrate multi-source data, including points of interest (POIs), street view images, and social media data. Taking Nanjing, China, as a case, we identified 192 RPs and their perceptual ranges (PRRPs). The results show the following: (1) Comparing RPs to non-RPs, RPs show higher average scores across four perceptual dimensions (positive indicators): Beautiful (7.11% higher), Lively (34.23% higher), Safety (28.42% higher), and Wealthy (28.26% higher). Conversely, RPs exhibit lower average scores in two perceptual dimensions (negative indicators): Boring (79.04% lower) and Depressing (20.35% lower). (2) Across various perceptual dimensions, RPs have utilized 15.13% of the land area to effectively cover approximately 50% of human perceptual hotspots and cold spots. (3) The RPs exhibit significant variations across different types, levels, and human preferences. These results demonstrate the positive perceived effects that RPs have, providing valuable insights to support urban management, the transformation of the built environment, and the promotion of sustainable urban development, and provide guidance for urban planners and designers to make improvements in urban design and planning to make these sites more attractive.

1. Introduction

Rapid urbanization in recent decades has led to a significant increase in urban populations, generating substantial demand for ecological, recreational, and shopping experiences [1,2]. Places in cities that provide ecological, recreational, and shopping functions are usually well developed [3], either dense and vibrant with commercial and public facilities, attracting a large number of human activities [4,5], or offering comfortable ecological spaces that allow for access to nature, promoting residents’ physical health [6] and mental health, and improving their sense of wellbeing [7]. These places often possess the unique character and charm of the city and are highly representative of it [8,9]. We can call these Representative places (RPs).

Representative places (RPs) is not a concept widely recognized by the academic community. Kevin Lynch lists five features for identifying cities in his work Urban Imagery, including nodes, landmarks, edges, paths, and zones [10]. The city can be represented and viewed through popular locations within these elements [11,12]. Therefore, Representative places (RPs) can be surface-like spaces such as commercial districts and tourist areas, linear spaces such as characteristic commercial streets and riverfront scenic zones, and point-like locations such as landmarks and urban nodes. However, this functional richness, as well as spatial diversity, makes RPs a collection of multiple research objects, posing a challenge to RP research. Nevertheless, since RPs play an important role in enhancing the local cultural perceptions of tourists and improving the life experiences of citizens [13], it is crucial to focus on RPs.

Identifying RPs and defining their spatial boundaries has become one of the primary challenges in the study of RPs, as urban research, urban management, environmental restoration, planning, and design all require a clear understanding of an object’s extent. Existing place identification studies have utilized appropriate data sources and effective tools to perform identification from various perspectives. To enrich the semantics of places, researchers have used point-of-interest and social media data to identify places and their functions and types. This is usually achieved by analyzing density or quantity and is synchronized with the identification of the places’ scope [14,15]. Social media platforms such as Twitter, Weibo, and Flickr provide many explicit geographic coordinate points that are generated by the public and are semantically rich [16,17]. However, these coordinate points often suffer from location bias, as users usually prefer famous attractions or popular places for check-ins, even if they are not actually there [18]. To obtain more accurate check-in locations, some studies have employed data-assisted methods. For instance, to account for location bias, J. Huang et al. [19] concatenated geotagged text and image data from Twitter and Instagram with survey data. POI data are the second most commonly used data source. Because it carries rich functional semantics, researchers use semantic information to identify the functional space of places. For example, Pan et al. [20] used POI data to identify the neighborhood functional areas of Macau Peninsula, Yang et al. [21] delineated the functional areas of the city by stacking buildings with POI data, and K. Liu et al. [14] identified the cognitive boundaries of transportation nodes using POI data. It is evident that most of these studies apply data to large-scale analyses of faceted urban functional areas, with less focus on the point and line spaces of small-scale places. In terms of spatial boundary identification methods, polygonal algorithms [22] and kernel density estimations [23] are commonly used to identify the ranges. Hobel et al. [24] employ natural language processing and machine learning techniques for identification, while K. Liu et al. [14] utilize DBSCAN clustering and Chi-shape algorithms, among other techniques, to generate cognitive boundaries for metro stations. Although these studies make useful attempts to address boundary ambiguities, they do not effectively address perceptually based boundary recognition. Furthermore, even when boundary ranges are extracted, issues such as lack of accuracy and irregular shapes persist.

In summary, to accurately identify places and delineate the scope of place perception, it is essential to select data that accurately reflects human perception of places and to find innovative ways to associate people with places. Since RPs are multifactorial collections, there have been no direct studies on RPs; researchers have predominantly focused on environmental spaces. Numerous studies have examined human cognition and perception of the environment through the interaction of the sensory organs with the environment, encompassing five aspects: hearing, touch, smell, taste, and vision. Auditory research focuses on the urban acoustic environment, particularly on how urban sounds impact human psychology and emotion [25,26,27]. Research in haptics (thermal comfort) focuses on skin contact with thermal environments, particularly on how these environments affect human health risks [28,29]. Additionally, more studies are focusing on the visually impaired, examining their tactile perception of crosswalks [30,31]. The term “smellscape” is frequently used to describe how the sense of smell, spanning across spaces, significantly impacts human behavior, attitudes, and health [32]. However, it has rarely been addressed or measured in planning practice and theory so far. Recent efforts have focused on understanding and regulating environmental odors, particularly on detecting, controlling, and eliminating maladaptive odors [33,34]. Taste research in urban environments is also relatively scarce at present. Existing studies primarily focus on the types of food offered by restaurants and ordered by users, examining the distribution of tastes in these environments [35,36] and the taste behaviors of the population [37]. Vision is the primary focus of built environment studies, which are mainly conducted through visual images, such as streetscape pictures, green landscapes [38,39], land use [40], and transportation spaces [41]. In particular, the development of machine learning and deep learning technologies has enabled large-scale, high-resolution, and multi-dimensional perception studies of built environments [42,43]. In summary, the research on built environments based on human perception is very large and fruitful. The relevant results are mostly from the visual perception level, and the research on hearing, touch, smell, taste, and so on is relatively small, which is mainly due to the difficulty of obtaining data sources. This is mainly due to the difficulty of obtaining data sources. So far, the acquisition of large-scale and high-granularity data sources has not been well solved. Existing visual perception research methods are relatively mature. Street view images are a location-based quality data source that provides a data base for place-based visual perception research. Deep learning-based human perception of street view images greatly facilitates visual perception quantitative research. Although visual perception research is more fruitful, less attention has been paid to comparative studies of specific places in cities, and there is also a lack of systematic visual perception evaluation of these places.

To address these issues, this study proposes a novel approach to identify and evaluate RPs (RPs) in cities from a human perception perspective. Our approach integrates multi-source data, including POIs, street view images, and social media, using techniques such as deep learning and text semantic analysis. Initially, popular check-in locations were extracted from microblog check-in data. Based on the matching pattern between these popular locations and POI names, a cohesion method that combines spatial clustering and polygon generation techniques was proposed to identify RPs and to extract the perceptual regions (PRRPs) of RPs. Subsequently, using the six dimensions of human perception as assessment indicators, the human perception characteristics of PRRPs were examined in terms of different types, levels, and preferences of various groups of people, and the perceptual differences between RPs and non-RPs were evaluated. We will address two key scientific questions: (1) How can RPs (RPs) and their Perceptual Ranges (PRRPs) be identified in terms of human perception? (2) How do human perceptions at RPs vary across types, levels, and population preferences? By answering these questions, this study aims to deepen our overall understanding of Representative places in the city, provide valuable insights into placemaking, management, and sustainability in the city, and provide guidelines for urban planners and designers in place planning and design.

2. Study Area and Datasets

2.1. Study Area

Nanjing, the capital of Jiangsu Province in China, is one of the most significant political, economic, and cultural hubs in the Yangtze River Delta region. It is situated between 31°14′–32°37′ N latitude and 118°22′–119°14′ E longitude, encompassing an area of approximately 6,587 square kilometers. This study focuses on the central urban area of Nanjing, comprising Jiangnan Main City and Jiangbei New Main City. The region includes the entire administrative areas of Gulou, Xuanwu, and Qinhuai, as well as parts of Jianye, Qixia, Yuhuatai, Pukou, Liuhe, and Jiangning, covering an area of 804 square kilometers (Figure 1). Nanjing was selected as the study area for two key reasons. Firstly, Nanjing is abundant in both human and natural resources, which require evaluation. Secondly, the central region of Nanjing exhibits an imbalanced development pattern, with the southern part demonstrating greater strength than the northern part. Leveraging the advantages of the main city RPs south of the Yangtze River and enhancing the quality of the recently established main city RPs north of the Yangtze River are crucial steps towards achieving a balance between the north and the south.

2.2. Datasets

2.2.1. Social Media Check-Ins

Because social media data encompasses text, images, and geotagging, it is believed to represent spatial and temporal information about human activity [44]. In this study, we use Sina Weibo user check-in data and data from the first ten pages of tweets posted by these users on their homepage. Sina Weibo is one of the most popular social media platforms in China, with an active user base of nearly 500 million, and is often referred to as the “Chinese Twitter” [15]. We developed a Python-based crawler program to collect microblog check-in data from Nanjing for the entire year from January 2023 to December 2023, resulting in 302,712 entries involving 157,461 users. Additionally, we collected the homepage data of these users, totaling 13.8 million entries. In addition to standard information such as user ID, posting time, and body content, each check-in record includes the name of the user’s check-in location, as well as the latitude and longitude coordinates of the check-in.

2.2.2. Points of Interest

POI stands for “Points of Interest”, which refers to geo-location data providing information about the characteristics and functions of places within a city. POI contains rich semantic information and is well suited for studies related to urban functions [14]. The POI data used in this study were obtained from the AMAP API service interface. All POI data for the year 2023 were collected in bulk using Python 3.11-based web crawler software. There are a total of 23 broad categories of POIs in AMAP. We selected 22 categories except residential, resulting in a total of 742,622 records.

2.2.3. Place Pulse 2.0

“Place Pulse 2.0” is a crowdsourcing platform for cityscape ratings [45] that collects preferences for paired images from online volunteers by showing them paired images, such as “Which place looks safer?”. This study investigates public impressions of six dimensions, ‘Safe’, ‘Beautiful’, ‘Depressing’, ‘Lively’, ‘Affluent’, and ‘Boring’, utilizing scores rather than pairwise comparisons, following the methodology of Salesses et al. [46] and Zhang et al. [43]. Specifically, for each perception, the win rate ($W_i$) and loss rate ($L_i$) of image $i$ are defined as:

$$W_i={\displaystyle \frac{w_i}{w_i+l_i+t_i}};L_i={\displaystyle \frac{l_i}{w_i+l_i+t_i}}$$

(1)

where $W_i$, $L_i$, ${\mathrm{w}}_i$, ${\mathrm{l}}_i$, and ${\mathrm{t}}_i$ represent the number of times image $i$ has won, loss, or equals with the paired image, respectively. The Q-score for each image $i$ ranges from 0 to 10 and is defined as follows:

$$Q_i={\displaystyle \frac{10}{3}}(W_i+{\displaystyle \frac{1}{w_i}}\sum _{j_1=1}^{w_i}{W}_{j1}-{\displaystyle \frac{1}{l_i}}\sum _{j_2=1}^{l_i}{L}_{j2}+1)$$

(2)

where $j_1$ and $j_2$ denote the images that loss or won over image iii in the comparison. Equation (2) corrects the win rate ($W_i$) of an image by adding the average win rate of the selected image and subtracting the loss rate of the selected image [46]. Thus, the Q-score incorporates the image information paired with each image [43,46].

2.2.4. Street View Images of Nanjing

Street view images of the study area were retrieved from the Baidu platform via a static map API (https://lbsyun.baidu.com/). The platform has been archiving street view images since 2013, with many updated in a timely manner. We sampled one street view every 100 m based on the road network and acquired a total of 323,560 street view images of Nanjing. The size of each street view image was fixed at 400 × 300 pixels, consistent with the Place Pulse 2.0 dataset.

3. Methodology

The framework of this study is illustrated in Figure 2 and comprises three phases: identification of RPs, human perceptions and evaluation of RPs.

3.1. Identifying the RPs

The number of tweets characterizes the popularity of a place [47], and tweet check-in locations filtered by a certain number of tweets are used as candidate RPs. Then, the perceptual ranges of the RPs are determined, and the candidate RPs that successfully obtain these perceptual ranges are taken as the final RPs. In this study, we propose using POIs to determine the perceptual ranges of RPs. This is because, when people name various facilities and locations, they associate them with distinctive RPs to make them memorable. For example, Nanjing Fuzimiao is a well-known attraction, and some service facilities are usually associated with this attraction when they are named, such as Hanting Hotel Fuzimiao Store, Duck Different Saltwater Duck (Fuzimiao Store), and JieDian (Fuzimiao Store). Different types of service facilities such as accommodation, catering, and electronics sales are included here. K. Liu et al. [14], Pan et al. [20], and others also point out that POIs can be applied to the study of place perception, cognitive mapping, and spatial behavior. Specifically, we retrieve POIs by the name of the RP, and the distribution of POIs is used to represent the scope of perception. Drawing on the Chi-shape algorithm proposed by Duckham et al. [48] and the research methods of Akdag et al. [49] and K. Liu et al. [14], we use the optimal set of concave hulls to determine the perceptual range of an RP.

The optimal point set concave hull is obtained using the Chi-shape algorithm. The Chi-shape algorithm generates a concave body by constructing Delaunay triangles and sequentially deleting initial boundary edges based on edge length until the longest remaining edge is shorter than a specified threshold. In the study by Akdag et al. [49], the lengths of the edges are normalized to [1, 100] and controlled by the parameter ${\lambda }_p$. Different values of ${\lambda }_p$ generate different concave bodies. The optimal value of ${\lambda }_p$ is determined by minimizing the fitness function to generate concave bodies with balanced complexity and hollowness. The fitness function comprises a concave body P, a set of Delaunay triangles D, and a parameter C that balances complexity and vacuity:

$$\varphi \left(P,D\right)=Emptiness\left(P,D\right)+C*Complexity\left(P\right)$$

(3)

where C is in the range [0, 1], and the larger the value of C, the less complex the polygon. In conjunction with Liu (Metro Cognition) et al. [14], this study sets C = 1 to obtain smoother concave bodies. Different values of ${\lambda }_p$ correspond to different fitness scores ϕ, and the value of ${\lambda }_p$ that corresponds to the smallest ϕ is the best parameter.

K. Liu et al. [14] applied the optimal point set concave hull method to cognitive range acquisition at subway stops. However, the resulting concave hull suffers from two issues: narrow band borders that should not be used as a cognitive range, and sharp vertices that detract from esthetics. To this end, we propose improved methods for concave hull acquisition. First, we cluster POIs using the DBSCAN method to circumvent band boundaries. Second, the concave hull is constructed using the four vertices of the mesh to avoid producing sharp vertices (Figure 3).

3.2. Classifying the RPs

We propose to identify the type of PR by comparing cosine similarity, which involves three main steps.

The first step is to train the word vector model.

We categorized RPs into Parks, Streets, and Squares [50], and then obtained textual introductions for these three classifications from Baidu Encyclopedia and Wikipedia. These texts, along with microblog check-in data, are used as a corpus for word vector model training. Finally, the CBOW (Continuous Bag of Words) model was chosen to train the word vectors [51]. The CBOW model brings semantically similar words closer in the vector space by learning a vectorized representation of each word. During training, the input is the word vector in the context and the output is the word vector of the target word (Figure 4). The CBOW model is trained on a large amount of textual data and learns a distributed representation (word vectors) of each word such that semantically similar words are close together in the vector space. An advantage of the CBOW model over other neural network models is that it trains relatively quickly and performs well with large amounts of data. We use the Python 3.11 package Gensim for word vector model training. The word vector size is set to 100, meaning a 100-dimensional vector represents each word. The window size is set to 10, meaning 10 words on either side of the target word are used to predict it.

The second step is to get the denotative word for place.

Before using the word vector model, we obtain the representative words for Parks, Streets, Squares, and all RPs. We use TF-IDF (Term Frequency–Inverse Document Frequency) to extract the 20 most important terms (

Table 1. Sample expressions for the type of place and the name of the place.

Type of Premises/Name of Premises	Indicators (Partial)
Park	Parks, green spaces, facilities, landscape, recreation, environment, science, places, sites, functions…
Street	Pedestrian street, leisure, pedestrian, urban, commercial, environment, shopping street, tourism, shopping, interior, landscape…
Square	Square, space, enclosure, design, architecture, town hall, environment, activity, building, pedestrian flow…
Xinjiekou	Eat and drink, shopping, nail, yummy, restaurant, vermicelli soup, gourmet, square, milk tea, tea yen…
Xuanwu Lake Park	Sunset, cherry blossoms, running, traveling, spring, lake, park, photography, weekend, spring…
D9 Street	Neighborhoods, private rooms, decks, photography, bouncing, lamb, romance, parties, drinking enough, weekend…

). TF-IDF is a common technique used in information retrieval and text mining to assess the importance of a word within a document relative to a document set or corpus [52]. The TF-IDF is calculated as:

$$TF-IDF = TF(t,d)\times log{\displaystyle \frac{N}{n_t}}$$

(4)

where TF(t,d) denotes the term frequency of word t in document d, and N is the total number of documents in the corpus and is the number of documents containing word ${\mathrm{n}}_{\mathrm{t}}$.

The third step is to obtain the classification of RPs using cosine similarity.

After extracting the representative words, the word vector model is used to generate the representative vectors. Then, cosine similarity is used to compare the similarity of each PR with Parks, Streets, and Squares, leading to PR classification. Cosine similarity measures the degree of similarity between two non-zero vectors, taking values between −1 and 1, with 1 indicating complete similarity, 0 indicating complete dissimilarity, and −1 indicating complete uncorrelation and completely opposite directions. The cosine similarity is calculated as follows:

$$\cos \left(\theta \right)={\displaystyle \frac{A·{\rm B}}{‖A‖‖{\rm B}‖}}$$

(5)

where A∙Β is the dot product of the vectors A and B, and $‖\mathsf{A}‖$ and $‖\mathsf{{\rm B}}‖$ are the Euclidean norms of the vectors A and B, respectively.

The word vectors of each RP are compared with the word vectors of Parks, Streets, and Squares for cosine similarity. The RP is assigned to the category with the highest cosine similarity value.

3.3. Identifying the Crowds and Their Preferences

It is important to identify the preferences of tourists and local residents for RPs, as different groups have different aspirations. The method consists of the following two steps:

The first step is to identify the crowd type. Inspired by previous work [15,53], we identify whether a user is a local resident or a tourist based on their spatiotemporal behaviors in their tweet records over time. Using the NER model, we extract each user’s spatial location in chronological order and construct it as a sequence of locations. We determine whether a Weibo user is a tourist or a local resident based on the following two rules. Rule 1: If the number of location elements in the location sequence is at least 3 and “Nanjing” appears more than 50% of the time, the user is a local resident; otherwise, the user is a tourist. Rule 2: If “Nanjing” appears more than once in an interval within the de-emphasized location sequence, the user is a local resident; otherwise, the user is a tourist.

The second step is to identify the preferences of different populations for RPs. We have categorized RPs into four groups: visitors like and local resident dislike, local residents like and visitors dislike, both local residents and visitors dislike, and both local residents and visitors like. We propose the following method for type differentiation. First, two types of places are distinguished: those loved by local residents and those loved by visitors. If more than two-thirds of the tweets are checked in by local residents, it indicates that local residents love the place; if more than two-thirds of the tweets are checked in by visitors, it indicates that visitors love the place. Then, a distinction is made between the types that both local residents and visitors love to visit and those that neither local residents nor visitors love to visit. If neither local residents nor visitors tweeted more than two-thirds of their tweet check-ins, compare whether the number of tweet check-ins for that RP exceeds the median number of tweet check-ins for all RPs. If it does, it indicates that both local residents and visitors love to go there; otherwise, it indicates that neither group loves to go there. The formula can be expressed as:

$$P_i=\left\{\begin{array}{l}{V}_1{L}_0, if N_{vi}>{\displaystyle \frac{2}{3}}{N}_i \\ V_0{L}_1, if N_{li}>{\displaystyle \frac{2}{3}}{N}_i\\ \begin{array}{c}{V}_1{L}_1, if N_{vi}\le {\displaystyle \frac{2}{3}}{N}_i and N_{li}\le {\displaystyle \frac{2}{3}}{N}_i and N_i>N_{median}\\ V_0{L}_0, if N_{vi}\le {\displaystyle \frac{2}{3}}{N}_i and N_{li}\le {\displaystyle \frac{2}{3}}{N}_i and N_i\le N_{median}\end{array}\end{array}\right.$$

(6)

where ${\mathrm{P}}_i$ denotes the $i$th PR, ${\mathrm{V}}_1{\mathrm{L}}_0$ denotes that visitors like it and local residents do not, ${\mathrm{V}}_0{\mathrm{L}}_1$ denotes that visitors do not like it and local residents do not like it, ${\mathrm{V}}_1{\mathrm{L}}_1$ denotes that both visitors and local residents like it, and ${\mathrm{V}}_0{\mathrm{L}}_0$ denotes that both visitors and local residents do not like it; ${\mathrm{N}}_i$ denotes the number of microblogging check-ins in the ith PR, ${\mathrm{N}}_{\mathrm{v}i}$ denotes the number of microblogging check-ins in the i PR that are visitors, ${\mathrm{N}}_{\mathrm{l}i}$ denotes the number of microblogging check-ins in which are local residents, and ${\mathrm{N}}_{\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}}$ denotes the median number of all RPs.

3.4. Mapping Human Perceptions Using ResNet50

This study employs a deep learning approach to quantify human perception of urban landscapes by analyzing street view images, involving three main steps, as follows.

3.4.1. Dataset Construction

Drawing from Wei et al. [40], Zhang et al. [43], and others, we exclude images with median perceptual scores and select street scene images with explicit perceptual labels for the binary classification process. Specifically, by calculating the mean ($\overline{\mathrm{Q}}$) and standard deviation ($\mathsf{\sigma }$) of the perceptual scores, images above the threshold ${\mathrm{Q}}_{\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}}$ are labeled as positive, and images below the threshold ${\mathrm{Q}}_{\mathrm{l}\mathrm{o}\mathrm{w}}$ are labeled as negative:

$$Q_{high}=\overline{Q}+\sigma , Q_{low}=\overline{Q}-\sigma$$

(7)

We separated the photos into training and validation sets in a 4:1 ratio for model training and performance assessment.

Table 2. Numbers of images for training and validating.

	Label	Beautiful	Boring	Depressing	Lively	Safety	Wealthy
Train dataset	Positive	1406	422	842	2205	2630	1085
	Negative	1532	538	699	2080	2583	1175
Validation dataset	Positive	321	109	225	503	681	271
	Negative	414	132	161	569	623	294

shows the number of images used by the six perceptual models.

3.4.2. Model Training

ResNet50, a convolutional neural network (CNN) proposed by K. He et al. [54], was chosen as the deep learning model. During the model training process, the cross-entropy loss function (${\mathrm{L}}_{\mathrm{c}}$) was used to calculate the difference between the true value and the predicted value, and accuracy (${\mathrm{A}}_{\mathrm{c}\mathrm{c}}$) was used as the model performance evaluation metric, with the following formula:

$$L_c=-\sum _i^n{y}_i\log \widehat{y_i}$$

(8)

$$A_{cc}={\displaystyle \frac{N_{\widehat{y_i}=y_i}}{N}}$$

(9)

where $\widehat{y_i}$ denotes the predicted category, $y_i$ denotes the true category, n represents the total number of classifications, $N_{\widehat{y_i}=y_i}$is the number of images with correct classification results, and N is the total number of images.

3.4.3. Model Prediction

After the six models were trained, street view photos from the study region were classified and predicted. For prediction, we were given the probability of each image being categorized into different perceptual categories (Positive vs. Negative). The probability of each street scene image was multiplied by 10 and converted to a [0, 10] perception score, which reflects the extent to which the image was perceived as ‘safe’, ‘beautiful’, ‘depressing’, ‘lively’, ‘affluent’, and ‘boring’.

3.5. Evaluating the RPs

To accurately assess human perceptions of RPs, we used the Getis-Ord $G_i^{*}$ statistic to identify statistically significant hot and cold spots [55]. The Getis-Ord $G_i^{*}$ statistic provides a z-score (also known as a $G_i^{*}$ value) and a p-value. The generated z-scores and p-values indicate whether grids with high or low values are spatially clustered or dispersed. The relevant formulas are as follows:

$$G_i^{*}={\displaystyle \frac{\sum _{j=1}^n{w}_{i,j}{x}_j-\overline{X}\sum _{j=1}^n{w}_{i,j}}{S\sqrt{\frac{\left[n\sum _{j=1}^n{w}_{i,j}^2-{\left(\sum _{j=1}^n{w}_{i,j}\right)}^2\right]}{n-1}}}}$$

(10)

$$\overline{X}={\displaystyle \frac{\sum _{j=1}^n{x}_j}{n}}$$

(11)

$$S=\sqrt{{\displaystyle \frac{\sum _{j=1}^n{x}_j^2}{n}}}-{\left(\overline{X}\right)}^2$$

(12)

where $G_i^{*}$ is the statistical z-score, $x_j$ is the perceptual score of grid j, $w_{i,j}$ is the spatial weight between grids i and j, and n is the total number of grids.

To further depict the spatial distribution of the six perception indicators in RPs, we developed a 500 m wide square grid covering Nanjing’s core city. The value of each square grid represents the average human perception scores of the street view images within that grid. A 500 m grid unit is very suitable for displaying spatial scales at the municipal level, as it effectively shows local spatial details and represents global spatial patterns [56]. Based on Getis-Ord $G_i^{*}$ statistics, we assessed the characteristics of RPs and non-RPs in terms of different types, levels, and population preferences.

4. Results

4.1. Identification of the RPs

4.1.1. Descriptive Statistics

We screened check-in locations in the city center with more than 50 tweet check-ins. After excluding locations named after administrative districts or those in the same location but called by different names, we identified 192 RPs (Figure 5a) and their perceptual ranges. These RPs are primarily located in the main urban area of Gangnam, accounting for 80 percent of the total, indicating an imbalance in the spatial distribution of RPs. When analyzing the perceptual range of the 192 RPs (Figure 6a), more than 60 percent have a perceptual range of less than 50 hectares. The largest perceptual range is Xuanwu Lake, covering 928 hectares, while the smallest is Yangtze River Park, covering only 3.04 hectares. There are nine RPs with an area exceeding 500 hectares, most of which are important business districts, transport hubs, and well-known scenic spots in Nanjing.

Multiple perceptual ranges may exist at RPs due to constraints such as natural conditions. A total of 63 RPs, or more than 30 percent of the 192 RPs, had two or more perceptual ranges (Figure 6b). Fifteen RPs had more than five perceptual ranges. These areas are mostly universities and scenic spots. Figure 7 shows several typical RP perceptual ranges. Xinjiekou (Figure 7a) and the Olympic Sports Center (Figure 7b) have more concentrated POIs, forming a single perceptual range. Nanjing South Railway Station is a high-speed railway station and the most important external transportation hub in Nanjing, around which a large number of lodging and food and beverage facilities have been developed. These facilities are not necessarily near Nanjing South Station, but rather at a distance from it, considering operating costs and superior transportation accessibility (Figure 7c). Nanjing Confucian temple is an important scenic spot where the main perceptual range is obvious, but there are still some secondary perceptual ranges in the form of enclaves (Figure 7d).

4.1.2. Category of the RPs

Nanjing is dominated by parks, with 89 RP types at 47.6%, followed by streets with 71 at 37.3%, and to a lesser extent, squares with 32 at 15.1%. In terms of spatial distribution of RPs (Figure 8a), representative parks types are widely distributed, with several concentrated areas, including Jiulong Lake Park, Fangshan, Xuanwu Lake, Xianlin University Town, Ming Xiaoling-Zhongshan Mausoleum, and the surroundings of the Olympic Sports Centre. In addition to traditional urban parks, the Park type also includes universities. This is because Chinese universities cover large areas, have good greenery, and are closed and managed. Microbloggers tweeting about campus environments on university campuses are more inclined to use words related to parks. The distribution of Street-type RPs is more concentrated, mainly located in the main urban area, including Nanjing Station-Maigaoqiao, Xinjiekou, Ming Xiaoling Mausoleum-Zhongshan Mausoleum, the periphery of Nanjing Software Park, and along the Qinhuai River (Figure 8b). Square-type RPs are spatially dispersed, including Xinjiekou, Nanjing South Railway Station, Nanjing Railway Station, and areas such as Hongyang Square–Jinxiang Shopping Centre in Pukou (Figure 8c).

4.1.3. Level of the RPs

In the context of urban regeneration or place transformation, planning needs to consider the importance of the place, making it necessary to grade RPs. The number of check-ins by Weibo users represents, to some extent, people’s recognition of these places. We categorized the number of microblog check-ins into three tiers by place type using the natural breakpoint method, resulting in Tier 1, Tier 2, and Tier 3 places based on check-in counts. There were four Tier 1 park places, five Tier 1 street places, and two Tier 1 square places, totaling eleven Tier 1 places. These places are some of the most recognizable locations in Nanjing and serve as windows to the city. For example, Nanjing Olympic Sports Centre is the most important sports event place in Nanjing and was the main place of the 2nd Youth Olympic Games in 2014. Nanjing Fuzimiao, built in 337 AD, is one of the largest traditional ancient markets in China. Nanjing Xinjiekou is a famous commercial center in China, with a history of over 100 years, and is regarded as “China’s first shopping district”. In addition, we have identified 27 secondary places and 187 tertiary places, which, together with 11 primary places, constitute the most distinctive areas of Nanjing and represent the overall image of the city.

We further examined the spatial distribution of RPs at all levels. As shown in Figure 9a, the Tier 1 places exhibit a remarkable belt-like layout, starting from Hongshan Zoo in the north, passing through Xuanwu Lake around Central Road and Zhongshan Road, and extending south to Xinjiekou and Fuzimiao. Additionally, Xuanwu Lake, Xinjiekou, Fuzimiao, and Nanjing Olympic Sports Center all have large perceptual ranges, while Niushoushan has multiple scattered perceptual ranges due to more dispersed POIs. Compared to Tier 1 places, Tier 2 places (Figure 9b) and Tier 3 places (Figure 8c) are perceived to be smaller in extent and more sporadically distributed.

4.1.4. Popularity of the RPs

The popularity of each RP was further analyzed based on the identified local residents and visitors. We identified areas that visitors love to visit but local residents do not (Figure 10a), such as Maigaoqiao, Tiesinqiao, Yanziji, Zhongguangmen, Chaotiangong, and Jixingmen. These places are intra-city transportation node areas where visitors tend to check in. For local residents, these areas are merely daily commuting zones, to which they do not pay special attention and where they do not stay for long. We also found some places that local residents love to visit but visitors do not (Figure 10b), such as Hongshan Zoo, Ancient Jiming Temple, Nanjing Museum, Niushoushan, Nanjing Olympic Sports Centre, and Nanjing Deji square. This is logical, as Nanjing is an ancient historical capital of China with a large number of well-known attractions, and lesser-known and locally serviced places such as zoos and sports centers are less visited by visitors. We also found some places that are not favored by either local residents or visitors (Figure 10c). These include Nanjing’s second-tier parks such as Egret Island Park, Qingliang Mountain, Linggu Scenic Spot, and Yueya Lake Park, as well as lesser-known universities and colleges like Nanjing University of Finance and Economics, Nanjing No. 2 Teachers’ Training College, Nanjing Information Vocational College, and Nanjing Railway Institute of Vocational and Technical Studies. Additionally, second-tier business districts such as Zifeng Mansion, Jinmao Hui, and Mingfa Commercial square, along with non-traditional commercial streets like Shanghai Road, Mochou Road, and Shanyin Road, are also less frequented by visitors and local residents alike. The major business districts and scenic spots in the central city, such as Xuanwu Lake Park, Xinjiekou, Fuzimiao, and Zhongshan Scenic Spot, are places that both visitors and local residents love to visit (Figure 10d). Nanjing’s major transport hubs, such as Nanjing South Railway Station and Lukou International Airport, are also popular with both visitors and local residents.

4.2. Evaluation of the RPs

4.2.1. Performance of the Deep Learning Model

This study trained six perceptual models using ResNet50. The model and the entire computation were implemented on the PyTorch platform (1.13.1+cu117), utilizing a single GeForce RTX4090 GPU. The optimization algorithm used was Adam, with the learning rate decayed on the 5th, 15th, and 20th epochs by a decay factor of 0.1. The dataset consisted of 20 images per batch, and each model was trained for 50 epochs. Our models were evaluated on the validation dataset, and all achieved an accuracy of more than 90%, with the Safety model showing the best results, converging at the 21st epoch and reaching an accuracy of 91.5% (Figure 11). The performance of our models is consistent with the findings of Wei et al. [40] in their study of human perceptions of urban landscapes in Shanghai.

4.2.2. Overall Characteristics

Figure 12 illustrates the perception scores of the six dimensions in the central city of Nanjing, showing typical spatial differentiation characteristics. We performed a mean aggregation of the perceptual scores of the 192 RPs and compared them with the perceptual scores of the non-RPs (Figure 13). The mean and median of the Positive index (Beautiful, Lively, Safety, Wealthy) of the RPs are greater than 5, while the mean and median of the Negative index (Boring, Depressing) are less than 5, reflecting positive attitudes towards urban space. All of the perception indicators of the RPs are better than those of the non-RPs. Specifically, the mean values of ‘Beautiful’, ‘Lively’, ‘Safety’, and ‘Wealthy’ are higher than those of the non-RPs by 7.11%, 34.23%, 28.42%, and 28.26%, respectively, while the mean values of ‘Boring’ and ‘Depressing’ are lower than those of the non-RPs by 79.04% and 20.35%, respectively.

Table 3. PRRPs cold hotspot grids as a proportion of all cold hotspot grids in the central city.

Spot Types	Confidence	Beautiful	Boring	Depressing	Lively	Safety	Wealthy
Hot spot	$>99\mathrm{\%}$	20.71%	0.49%	3.65%	40.52%	31.67%	31.18%
	$>95\mathrm{\%}$	36.28%	1.96%	8.21%	52.32%	45.79%	47.56%
	$>90\mathrm{\%}$	47.08%	2.94%	10.18%	56.75%	53.20%	55.04%
Cold spot	$>99\mathrm{\%}$	5.23%	44.38%	28.40%	1.82%	2.15%	2.80%
	$>95\mathrm{\%}$	15.12%	53.00%	44.61%	3.18%	4.76%	6.05%
	$>90\mathrm{\%}$	23.72%	57.55%	55.99%	4.55%	7.68%	7.96%

shows the proportion of cold hotspot grids with RP-aware ranges (PRRPs) to all cold hotspot grids in the central city. Higher values indicate more highly perceived grids of PRRPs, which are spatially clustered areas of high values. It can be seen that the percentage of hot spot grids for the positive indicators (Beautiful, Lively, Safety, Wealthy) and the percentage of cold spot grids for the negative indicators (Boring, Depressing) are higher than the percentage of the area of the PRRPs relative to the area of the central city limits across all three confidence intervals (15.13%). At the 90% confidence level, the positive indicators all have a hot spot coverage of around 50%, while the inverse indicators all have a cold spot coverage of more than 50%. This suggests that RPs concentrate most of the hot (positive indicator) or cold (negative indicator) areas, further emphasizing the positive urban perception effect of PRRPs.

The spatial distribution of cold spots and hot spots grids shows typical heterogeneity. As shown in Figure 14a, the Beautiful perception indicator is generally lower in areas outside the PRRPs around Zijinshan Mountain and Xianlin University Town. In the vicinity of Nanjing South Railway Station, significant cold spots are observed. Other than that, the vast majority of the PRRPs are neither hot nor cold spots. For the Boring perception dimension (Figure 14b), there is a large range of cold spots centered on Xinjiekou. The vast majority of PRRPs are significantly cold, except for a small number of PRRPs northwest of Nanjing South Station and Xianlin University Town. For the Depressing perception dimension (Figure 14c), Nanjing South Station and Nanjing Station presented significant hot spots, while the rest of the PRRPs exhibited obvious cold spot characteristics. However, the PRRPs in the downtown area and the Maigaoqiao area did not show significant hot spot or cold spot characteristics. Lively and Boring show opposite spatial characteristics (Figure 14d), with Boring’s cold spots coinciding with Lively’s hot spots. Notably, none of the PRRPs are significantly cold. In addition, we can also see that compared to Safety and Wealthy, Lively is more widely distributed and has a stronger degree of aggregation, especially in the main city of Jiangnan, which presents a stronger centralized contiguous feature. The spatial characteristics of the Safety and Wealthy perceptual dimensions are similar, with the vast majority of PRRPs being significant hot spot areas. The difference is that Nanjing Station’s Safety perception shows cold spot areas, whereas in Wealthy there are no significant hot or cold spots. In general, in the Positive index, PRRPs are mostly hotspot areas except Beautiful, and areas such as motorways, urban expressways, and large bridges are mostly cold spot areas; in the Negative index, the overall presentation is opposite to the characteristics of the Positive index, with areas such as motorways, urban expressways, and large bridges mostly being hotspot areas, and PRRPs mostly being cold spot areas.

4.2.3. Evaluation of Different Types

Table 4. Human perception scores for different classifications of PRRPs.

Areas	RP Types	Beautiful		Boring		Depressing		Lively		Safety		Wealthy
		Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median
RP-wide	Park	5.346	5.390	2.869	2.720	4.122	3.910	7.183	7.460	6.162	6.390	5.087	5.160
	Street	4.580	4.580	2.160	1.700	4.891	4.890	7.758	8.310	6.225	6.560	4.900	4.950
	Square	4.979	4.930	2.092	1.770	4.357	4.230	7.687	8.080	6.038	6.340	4.870	4.980
Non RP-wide	--	4.878	4.870	4.263	4.500	5.026	4.990	5.696	5.550	5.000	4.800	4.044	3.900

demonstrates the human perception scores for the different classifications of PRRPs, and Figure 15 shows the percentage of cold and hot spots for the different classifications of PRRPs for each type of grid. We found that parks appeared relatively Beautiful (Figure 15a), with a perceived mean of 5.346 and a hot spot grid percentage of 31.67%, both higher than the Street and Square. This can be attributed to the fact that green spaces may help parks appear more beautiful. However, for Boring (Figure 15b), parks are perceived as more Boring, with a mean score of 2.869 and a cold spot grid percentage of 64.04%, significantly lower than Street and Square. This is logical, as Street and Square typically offer a greater variety of services and a wider range of building types. Therefore, Parks are more Boring than the Street and Square. For the inverse indicator, Parks have a significantly greater percentage of Depressing cold spot grids than Street and Square (Figure 15c), suggesting that parks are less Depressing compared to the Street and Square. This logical comparison is likely due to Parks providing more green open space. For the Lively, Safety, and Wealthy dimensions (Figure 15d–f), PRRPs all exhibit a high percentage of hot spot grids, with squares having the highest percentage, characterizing their strong positive spatial effect.

4.2.4. Evaluation of Different Levels

Table 5. Human perception scores for different levels of PRRPs.

Areas	RP Levels	Beautiful		Boring		Depressing		Lively		Safety		Wealthy
		Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median
RP-wide	Level 1	4.645	4.600	2.212	1.690	4.866	4.810	7.470	8.020	5.894	6.170	4.689	4.720
	Level 2	5.542	5.520	2.715	2.560	4.120	3.850	7.189	7.460	6.072	6.260	5.077	5.200
	Level 3	5.127	5.150	2.492	2.250	4.199	4.040	7.536	7.850	6.230	6.500	5.058	5.150
Non-RP-wide	--	4.878	4.870	4.263	4.500	5.026	4.990	5.696	5.550	5.000	4.800	4.044	3.900

demonstrates the human perception scores for the different levels of PRRPs, and Figure 16 shows the percentage of cold and hot spots in the different classifications of PRRPs for each grid level. We found that the perception scores and hot spot grid percentages were significantly lower for Level 1 than for Level 2 and Level 3 in terms of Beautiful, Safety, and Wealthy perceptions. This is clearly inconsistent with common sense, as there is a tendency to think that the higher the rank, the stronger the positive effect the place should have. Our study confirms that higher levels of RPs do not imply stronger perceived positive effects. For Nanjing, Level 1 PRRPs can be further improved in the Beautiful, Safety, and Wealthy dimensions. The Boring perception score for Level 1 is 2.212, and the percentage of cold spot grids to grids at that level is 82.73%, indicating that Level 1 is less Boring compared to Level 2 and Level 3. The Depressing perception score for Level 1 is 4.866, and the percentage of cold spot grids at that level is 40.56%, indicating that Level 1 is more Depressing compared to Level 2 and Level 3. The Lively perception score of Level 1 is 7.470, and the percentage of hot spot grids to the grids of this level is 77.91%. However, there is not much difference between Level 2 and Level 3 in terms of scores and percentages, indicating that the Lively perception of Level 1 is not significantly different from that of Level 2 and Level 3.

4.2.5. Evaluation of Different People’s Preferences

Table 6. Perceived scores of PRRPs for different population preferences.

Areas	RP Categories	Beautiful		Boring		Depressing		Lively		Safety		Wealthy
		Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median	Mean	Median
UCN-wide	T0L0	5.119	5.090	2.668	2.330	4.280	4.160	7.729	8.100	6.366	6.680	5.072	5.160
	T0L1	5.323	5.390	2.194	1.850	4.099	3.870	7.755	8.150	6.827	7.160	5.584	5.790
	T1L1	5.148	5.110	2.555	2.270	4.402	4.220	7.266	7.600	5.955	6.200	4.878	4.960
	T1L0	4.877	4.880	2.339	2.100	4.285	4.160	7.743	8.100	6.291	6.570	5.048	5.150
CD-wide	--	4.878	4.870	4.263	4.500	5.026	4.990	5.696	5.550	5.000	4.800	4.044	3.900

presents the human perception scores of PRRPs for different population preferences, and Figure 17 shows the percentage of PRRPs cold and hot spot grids in each categorized grid for different population preferences. We find that places favored by both visitors and local residents do not exhibit the highest aggregation characteristics. Instead, they have the lowest percentage of cold spots (inverse indicator) and hot spots (positive indicator) within the categorical grid for the four perceptual dimensions of Boring, Lively, Safety, and Wealthy. Places that are disliked by visitors but liked by local residents have the highest perceived scores for the positive indicators (Beautiful, Lively, Safety, Wealthy) and the lowest perceived scores for the negative indicators (Boring, Depressing). Both positive and negative indicators show the highest aggregation characteristics for their hot and cold spots. In the Lively dimension, the hot spot grid share of places disliked by visitors but liked by local residents is 91.57%. This phenomenon is most likely due to the fact that there are fewer such RPs in the central part of Nanjing, leading to a statistical bias that does not reflect a general pattern. However, this conclusion is applicable to the Nanjing case. We also find that places disliked by both visitors and local residents do not exhibit the lowest aggregation characteristics. Instead, they score higher than places liked by both visitors and local residents as well as those liked by visitors but disliked by local residents on all three dimensions of the positive indicators (Lively, Safety, and Wealthy) and on the inverse indicator, Depressing.

5. Discussion and Conclusions

5.1. The RPs and Their Perceptual Regions

This study introduces an innovative approach to identify RPs and their perceived areas within the city through POIs combined with social media data. Our results identify 192 RPs in Nanjing. Of these, 89 are Parks, 71 are Streets, and 32 are Squares. Park-type RPs were the most common. On the one hand, Nanjing is surrounded by mountains and water, and the abundant natural resources within its borders reflect the characteristics of the city of Nanjing. On the other hand, it aligns well with the development intentions of the Nanjing Municipal Government. Nanjing proposes that by 2035, the coverage rate of green areas and public spaces larger than 400 square meters within a five-minute walking distance in built-up areas will reach more than 95%. With the significant increase in urban parks and green spaces, human perceptions of space tend to favor the Park type.

We categorized the RPs into three levels: 11 level 1 places, 27 level 2 places, and 120 level 3 places. The first level consists of highly recognized places that serve as windows to the city of Nanjing, but all of these places are located in the main city of Jiangnan. Additionally, the number of secondary and tertiary places in Jiangbei is far less than that in Jiangnan Main City. The Nanjing Territorial Spatial Master Plan (2021–2035) (Draft) proposes enhancing the functions and quality of the new main city in Jiangbei. Therefore, in the future, Nanjing needs to select qualified places for cultivation among the secondary and tertiary places in the main city of Jiangbei. This will help improve the functions of the new main city of Jiangbei by adding one or two primary places and several secondary places.

We identified four categories of places: visitors like and local residents dislike, local residents like and visitors dislike, both local residents and visitors dislike, and both local residents and visitors like. The number of places in each category was 11, 12, 13, and 22, respectively. In addition to the findings from the conclusions, we found that places popular with both visitors and local residents usually have a larger range of place perceptions. It is logical that the more popular an RP is, the more likely it is that the various facilities around it will be named after the RP, making it more memorable and recognizable. These findings can provide insights for subsequent RP remodeling and planning.

5.2. Human Perception of the RPs

RPs can be the city’s transport hubs, business districts, or well-known attractions, and providing high-quality RPs is crucial to enhancing the city’s image. We used six dimensions of human perception as assessment metrics for evaluating RPs. Our study reveals spatial differences in the perception of RPs. Overall, each perception indicator in the RP perception area shows better performance than in the non-RP perception area, indicating positive spatial effects within the RP perception area. This finding fully reflects the inherent representativeness of these places. For RPs with a small percentage of hot and cold spot grids, spatial optimization can be performed in conjunction with their area and identification. Previous studies have discussed the relevance of the six dimensions to streetscape elements. Therefore, the human perception score can be improved by adding more relevant streetscape elements.

5.3. Potential Applications and Research Limitations

RPs have the city’s unique personality and charm and best represent its character. Understanding the spatial distribution and human perception of RPs is essential to enhance visitors’ cultural perceptions and citizens’ life experiences. Firstly, RPs are identified by the popularity of visits, and the scope of the place is determined by people’s spatial perception, which truly reflects their willingness. Armed with this knowledge, urban planners and designers will be able to optimize or plan for different types and levels of places that meet the preferences of visitors and local residents, thus forming a comprehensive, functional, and hierarchical network of functions. Second, the human-perceived evaluation of RPs encompasses the human evaluation of the environment and facilities. Through these evaluations, it is possible to grasp people’s overall impression of the city at the macro level; at the meso level, it is possible to formulate guidelines for remodeling places of different types, grades, and preferences of different groups of people; and at the micro level, it is possible to put forward detailed remodeling suggestions for specific places. Compared to traditional small-scale analyses, the use of large amounts of social media and SVI data can cover the entire city and quickly complete the identification and assessment of RPs. This street-image-based approach offers advantages in terms of efficiency, cost-effectiveness, and time saving compared to a field survey of the same scale. However, it is important to acknowledge the limitations of this study. Firstly, social media data are limited to microblogging data, which is skewed and may exclude some population segments, resulting in biased place identification results [57]. For example, some older adults are also important populations for RP use, but they generally do not publish Weibo, so targeted supplementation of missing population data is necessary. Potential measures for this are increased questionnaires for missing groups and statistics on useful information about missing people in conjunction with attraction traffic data. Secondly, the static nature of the Baidu street view platform’s images hinders the acquisition of historical images, thus ignoring the impact of seasonal changes on human perception. In vegetation-rich places, there may be significant differences in landscape characteristics between winter and summer [58]. Therefore, methods such as crowdsourcing can be used to supplement missing and anachronistic streetscapes based on the time of the study. Thirdly, due to the limitation of the training dataset, we only used six dimensions of human perception as evaluation factors. Methods such as crowdsourcing can be used to add more targeted perceptual evaluation factors. Fourthly, the training dataset contains many European and North American cities, but not Nanjing, which could lead to potential perceptual biases due to changes in the urban environment. This can be accomplished by censoring out data with very different architectural styles and adding data for the target cities. Fifth, ResNet50 is not state-of-the-art, although it is a more effective neural network model. ResNet50 was chosen because of the low training data. Deeper, more advanced models require more data, a limitation that can be improved with subsequent increases in training data.

Due to space constraints, this paper does not discuss the influencing factors that form the spatial pattern of RPs. An in-depth discussion of the influencing factors can reveal Nanjing’s historical development and current economic pattern, thus providing more practical solutions for future urban development. In addition, based on the findings of the current RPs identification and assessment, a more complex analysis of the interactions between residents and tourists in the urban space can be carried out by combining the survey data or interviews, and the related findings can further support urban management, built environment renovation, as well as sustainable urban development. The above limitations can be gradually resolved in further studies in the future.