Characterizing and Measuring the Environmental Amenities of Urban Recreation Leisure Regions Based on Image and Text Fusion Perception: A Case Study of Nanjing, China

Abstract

Quantitative evaluation of the environmental amenities (EAs) in urban recreation and leisure regions (URLRs) can provide stronger support for the government to enhance the quality of urban leisure space and improve the well-being of urban residents. Considering the diversity of leisure spaces and the complexity of environmental perception perspectives, this study proposes a comprehensive environmental measurement framework based on image and text fusion perception, which utilizes big data to perceive and quantify the EA features of URLRs comprehensively and efficiently. The study of the URLRs in Nanjing, China, was conducted as an empirical study. The results indicate the following: (1) When it comes to leisure environments, the top concerns for most people are service, hygiene, reputation, and walkability. (2) The EA level of URLRs in Nanjing generally decreases from the center to the outside and shows regional differentiation. (3) EA features in Nanjing’s URLRs exhibit a spatial pattern of similarity in the center and at each district’s edges. This study enhances our understanding of leisure regions’ environmental features that contribute to quality. The measurement results support understanding the spatial heterogeneity patterns of urban leisure activities and vibrancy. Furthermore, valuable urban planning and policy suggestions are made to promote sustainable urban development.

1. Introduction

Environmental amenities (EAs) are an important aspect of urban livability, emphasizing the attractiveness of the environment [1] and the spiritual aspect of people [2,3,4]. A high level of EAs contributing to subjective well-being [5,6] can promote social welfare and economic gain [7]. Due to its close relationship with urban growth and sustainable development [8], EAs have been an essential factor in urban planning [9]. As one of the basic components of the urban built environment [10], the recreation and leisure regions provide the specified functionality of cities and improve the quality of life [11,12]. More attention has been given to the experience of users in urban recreation and leisure regions (URLRs), which impacts their likelihood to revisit, recommend, and share the venues [13,14,15] from an environmental psychological point of view. Therefore, the measurement of URLRs’ EAs can indicate the economic development potential and the underlying spatial heterogeneity patterns of urban leisure activities.

Environmental psychology provides an excellent theoretical basis for the study of perception between subjects and objects. A growing body of research has examined how the quality of environments, as perceived by humans, affects activities in URLRs. Specific realistic scenes, such as walking [16,17] and cycling [18], have been considered. Surveys such as questionnaires and interviews are the usual tools to obtain the environmental attributes and subjective perceptions of the interviewees [19,20,21,22]. Additionally, some scholars measure the EAs quantitatively by establishing a system of indicators and their weights scored by experts based on an analysis of the content or components of the study object [23,24,25,26]. The results revealed that perceptions of different environmental attributes have distinct relevance to people’s activities and which features are more significant. However, these methods are laborious, time-consuming, and costly for large study areas, and they might bring errors due to people’s biases. Besides, the urban environmental landscape changes rapidly, and people’s quest for environmental quality constantly changes and improves.

Crowdsourcing allows for the collection of urban perception data on a large scale. The proliferation of deep learning technology and computer vision has further enhanced the ability to perceive the urban environment [27,28,29]. Mining street view images’ semantic information can aid in comprehending and quantifying environmental features [30,31], which plays an essential role in landscape and urban planning. For example, some researchers use machine learning-based approaches to model human perceptions of the physical environment with street-level images [32,33]. Meanwhile, the quality of the environment, including the natural, human, and social aspects of the environment [34,35], can be measured by the sentiment analysis [36,37,38] of social media texts, such as Dianping (https://www.dianping.com, accessed on 25 October 2023) and Twitter (https://twitter.com, accessed on 25 October 2023). Texts express subjective perspectives, while images depict the physical environment. Both have strengths in portraying the environment, but they lead to differences in measurement accuracy.

Though evaluating the environment has accumulated a large amount of research by obtaining big data, a few studies have attempted to explore how to measure the EAs of the URLRs based on big data perception. Only some spaces have been studied in this regard. Streets, as fundamental elements in urban studies [39,40,41], not only have the function of transportation but also provide services for leisure activities [42,43]. For example, scholars use Google Street View images for semantic segmentation to measure the impact of street quality on recreational physical activities [44,45]. However, visitors also demand greater amenities regarding the indoor environmental quality for leisure [46,47], such as convenient facilities and air quality.

Thus, previous studies still need to adequately focus on a comprehensive evaluation of EAs for URLRs since a single piece of street view data or a description of an indoor space can only partially perceive the leisure environment. The rise of multimodal fusion technology [48] provides strong support for solving the above difficulties. Multimodal fusion refers to the association and reorganization of different modal data (such as text, image, audio, and video) at the data, feature, or semantic level based on deep learning to fully exploit the data’s value, thus solving complex urban problems. One benefit of combining data is an improved and integrated perception of EAs by utilizing the advantages of multiple sources. Another advantage is enhancing perception by combining descriptive features from different perspectives [49], creating semantic feature complementarity (spatio-temporal and other semantics).

To measure the EAs of URLRs more comprehensively and efficiently, referring to the multimodal fusion method, a novel framework has been developed in this paper using street images and web text fusion perception to depict the features of the EAs of URLRs and quantitatively evaluate the EAs of URLRs. The proposed framework based on EA theory and big data sensing is applied to analyze the characteristics of the spatial distribution of the EAs of URLRs in Nanjing, China. The innovations and contributions of this study are follows:

• A framework is proposed for measuring URLR environments using fused image and text perception. This considers leisure space diversity and complex environmental perception.
• The features of EAs in leisure spaces are explained by big data fusion perception, enriching the system of EAs and the connotation of EAs for leisure. The comprehensive and quantifying perception of EA features in URLRs provides a clearer pathway for leisure quality enhancement.
• The time and space distribution of EAs in URLRs of a developing country city is analyzed at the more refined and practical spatial scale, and its formation mechanism is preliminarily revealed. The specific guidelines for urban construction and leisure planning are presented based on the results of the study.

2. Data and Case Study Area

2.1. Data

2.1.1. Areas of Interest (AOIs) and Points of Interest (POIs)

According to the study [50], the recreation leisure pace covers amusement parks, shopping malls, shopping streets, supermarkets, SPA, barbershops, restaurants, bakeries, cold drink shops, bars, cafes, dessert shops, teahouses, KTV, gaming rooms, Internet cafes, nightclubs, urban squares, flower markets, home building materials markets, home appliance stores, and bath and massage houses. To obtain recreation leisure region, 267 AOIs of the above types were extracted from the AMAP (https://www.amap.com, accessed on 25 October 2023), one of China’s largest map service providers. Besides, 48,372 POI data of the above types were obtained from the AMAP, and they were clustered to form regions of 174 to supplement AOI data. POI data with intensive urban leisure function is also recreation leisure space. For example, the spatial scope of snack streets near residential communities provides the closest recreational area for nearby residents. In addition, POI data with a higher level of abstraction are widely used to detect the scope of urban functional areas [51,52,53]. They also have the potential to identify urban recreation leisure regions by using OPTICS (Ordering points to identify the clustering structure) and the Concave Hull method (Figure 1). The POIs and AOIs have the basic name, address, category, latitude, and longitude information.

2.1.2. Baidu Street Map View Image and Road Network

Among various multi-source data, street-level imagery portrays the visible urban environment from a human perspective. This indispensable tool [54] has allowed for perceptive analysis and comprehension of the built environment, inspiring the creation of innovative methodologies [33]. To assess the amenities of the street environment of the leisure area from the perspective of people’s subjective perception, 266,066 street view images of 1787 roads were obtained by Python programming from the Baidu Street View service (https://map.baidu.com, accessed on 25 October 2023), which has wide coverage and more timeliness. At an interval of 50 m, the sampling locations were generated along their urban road network in the study area. For each location point, the street view images cover four angles: 0, 90, 180, and 270 degrees, corresponding to the front, right, back, and left (Figure 2), and the vertical field of view of the image is set to 20 degrees.

2.1.3. Web Comment Text

Admittedly, the street view can describe the outdoor environment around the street. However, the street view cannot cover the internal space except the road. Additionally, the perception of the environment of leisure regions has connections not only to the objective of the environment but also to personal subjective preferences. Web comment text needs to be obtained for the study object. The Dianping website (https://www.dianping.com, accessed on 25 October 2023), China’s most popular lifestyle service review site, provides user reviews about various leisure and entertainment venues. Its huge hidden value can be mined, truly reflecting the EAs of leisure and entertainment. In this paper, we selected the Dianping website to capture the user comment data from 1 January 2017 to 31 December 2022. Over 1800,000 reviews for 441 URLRs were collected using Python programming. The detailed information of above data is shown in

Table 1. Summary of experimental data information.

Data	Content	Source	Time
POI	48,372 points’ coordinates and their attributes	Amap	April 2020
AOI	267 polygons with different functional attributes
BMSV images	82,808 street sampling points along with the transportation network	Baidu Maps
Roads	1787 urban roads
Web comment text	1,800,000 reviews with targets and date-times of 6 years	Dianping website	1 January 2017~ 31 December 2022

.

2.2. Study Area

As an important gateway city for the Yangtze River Delta region, Nanjing, the capital of Jiangsu Province, located in the middle and lower reaches of Eastern China (118°22′–119°14′ E and 31°14′–32°37′ N), was chosen as the case city. Nanjing has a long history and is at the forefront of leisure development in China. Nanjing’s 14th Five-Year Plan states that it will strive to become a happy and livable city with a higher leisure quality. Hence, it holds significant practical value in characterizing the quality of the leisure environment in this area. The city’s recreational spots and main residential areas are located mainly in the urban area. The study area consists of the primary urban area along with three suburban regions, as depicted in Figure 3.

3. Methods

3.1. Overview

The workflow of this study can be divided into three steps (Figure 4).

• In the first step, street view images and web comment texts about URLRs are used to extract features of EAs through objective and subjective perception, respectively. Then, the obtained features of each URLR are quantified by semantic segmentation and sentiment analysis, respectively.
• In the second step, to fetch the most influential EA features, the analytic hierarchy process (AHP) and entropy evaluation method (EEM) are used to determine the weights of the features. Additionally, a quantitative calculation is conducted to measure EA.
• In the third step, the spatial-temporal analysis is done based on quantitative results to explore the characteristics of the spatial and temporal patterns of EAs for Nanjing’s URLRs. Finally, urban planning and policy implications are proposed based on the results of this study.

3.2. Quantitative Measurement by Subjective Perception

Since quantifying subjective perception involves web text, and natural language processing (NLP) is a complex task, we use a flowchart (Figure 5) in this section to explicitly express the relationship between the steps, expressing how the EA subjective features are extracted and quantified.

3.2.1. Web Comment-Based EA Features Extraction

The user’s experience of the EAs of the URLR is implicitly or explicitly embedded in the text of the review. The clustering of high-frequency feature words has been used to extract features of EAs.

To capture aspects of people’s environmental concerns, high-frequency words related to environment description by sentence subdivision processing are enumerated. The lexical choices for these high-frequency words are nouns (n.), verbs (v.), adjectives (a.), and status words (z.) by the Jieba participle, which can characterize environmental features. Related high-frequency words are the primary feature words to extract the features of the EA. Additional words that are semantically similar to primary feature words are included (more process details can be seen in Appendix A). To further categorize the features of EA, the k-means clustering algorithm [55] is used to perform two-dimensional spatial clustering for the feature words. This clustering is based on the word embedding representations of the feature words using the Word2Vec model, which is good at implicitly learning relationships between words [56]. Additionally, the t-distributed stochastic neighbor embedding (T-SNE) [57] of their embeddings is needed to reduce the dimension of the embedding vector before the clustering. Examples of Word2Vec embedding visualization are shown in Figure 6.

3.2.2. Subjective Quantitative Measurement Based on Web Review Text

The ultimate aim is to quantify the EA features based on the characteristics of online text expressions. The degree of quantitative emotional values is used to characterize the level of EAs, following a two-step approach.

• In the first step, the review texts are categorized based on the EA feature categories. It is necessary to cut long texts into short texts for fine-grained sentiment measurement. To obtain information on the perceived EAs of the URLRs from a single brief comment text, the BERT model [58] is used to get short texts belonging to the category of EA features.
• In the next step, a combination of deep learning and sentiment lexicon-based methods is proposed for the quantitative analysis of sentiment values. Short comment texts may express positive or negative emotions with varying degrees of emotional expression and may not contain any emotional information. Bert+BiLSTM model is used for sentiment polarity classification, reaching 95% accuracy. Sentiment dictionary-based text analysis is used to quantify the value of emotional expression. Different levels of emotion can characterize different levels of EA. Referring to the relevant research [59], adverbs of the degree of emotion are classified into six classes for fine-grained emotional mining, namely “over”, “most/extremely”, “very”, “more”, “slightly”, and “insufficiently”. In addition, the exclamation mark and question mark can also characterize emotions. Different adjusted values are given to different grades of degree adverbs based on initial sentiment values according to emotional polarity. Examples of the adverb dictionaries and their adjusted values are shown in

  Table 2. Examples of the degree adverb dictionary.

  Grade	Word	Number	Adjust Value of Negative	Adjust Value of Positive
  over	a little over, far more than, unduly…	25	0.25	0.25
  Most/extremely	completely, unparalleled, extremely…	64	0.20	0.20
  very	very much, specifically, a lot, too…	42	0.15	0.15
  more	more and more, fully, all the more…	36	0.10	0.10
  slightly	a few, not so, mildly…	34	−0.10	−0.05
  insufficient	not really, a little less, not too…	13	−0.05	−0.10
  exclamation mark	!, !	2	0.05	0.05
  question mark	?, ?	2	1-value 1	1-value 1

  ¹ The value refers to the value calculated after the sentiment adverb is found.

  .

To ensure that the emotion quantification score remains within the acceptable range of 0 to 1, the negative emotion quantitative value must be below 0.5, and the positive emotion quantitative value must be above 0.5. We establish the initial positive emotion value at 0.7 and the initial negative emotion value at 0.3, according to the aforementioned adjustment values of the emotional adverbs. The formulas for calculating the sentiment value are shown in Equations (1) and (2). The basis and details of the specific reconciliation decisions can be seen in Appendix B.

value = 0.7 + adjust value

(1)

value = 0.3 − adjust value

(2)

3.3. Quantitative Measurement by the Objective Perception

3.3.1. Street View-Based EA Features Selection

The street is the main carrier for external recreational space and the main medium for circling activities within and outside the recreational area. The overall EAs of a street is shaped by how people visually perceive its physical environment. In this study, the EAs of street space are perceived through streetscape images from a human perspective. Drawing on previous research on the measurement environment of street images [31,60,61], four objective and static features of greenness (green space accessibility), openness, walkability, and enclosure are selected to complement the EA features of comment text, all impacting participation in leisure activities. Greenery and openness can enhance beauty and pleasure while reducing depression [62,63]. The enclosure provides a feeling of room-like space [64], promoting the perception of safety and comfort for visitors [63,65]. Walkability provides a strong perception of coherence, ensuring that pedestrian facilities are well-established [66].

3.3.2. Objective Quantitative Measurement Based on Street View Images

• Spatial connection between URLRs and spots of street view images

The condition of the street space and its surroundings greatly affects the overall EAs of a URLR. To ensure precision, we utilize ArcGIS 10.6 to spatially correlate the URLRs with the streetscape sampling points on the nearest roads in their vicinity. These streetscape points are generated every 50 m according to the Baidu map road network. Some URLRs are not linked to streetscape sampling points because they are far from the road. This suggests that the street environment in a few places does not impact the overall EA. Some of the spatial association examples are shown in Figure 7.

2.

Calculation of objective EA features

Semantic segmentation of street view images is necessary to perform quantitative perceptual calculation for street view images. Eleven kinds of physical components were obtained by DeepLab v3+ [67], namely “road”, “building”, “vegetation”, “pole”, “fence”, “sidewalk”, “trunk”, “sky”, “person”, “vehicle”, and others. The test results of semantic segmentation of the case street images can be seen in Figure 8.

Then, the percentage of vegetation as greenery (%), the sky ratio as openness (%), the proportion of vertical elements and horizontal elements as enclosure (%), and the proportion of road and sideways as walkability (%) were used, as shown in

Table 3. EA features for BMSV images and their calculation methods.

Features	Quantitative Methods
Greenness	$Greenness=\frac{{\sum }_{i=1}^n{vegetation}_i}{{\sum }_{i=1}^n{sum}_i}$
Openness	$Openness=\frac{{\sum }_{i=1}^n{sky}_i}{{\sum }_{i=1}^n{sum}_i}$
Walkability	$Walkability=\frac{{\sum }_{i=1}^n{sidewalk}_i+{\sum }_{i=1}^n{fence}_i}{{\sum }_{i=1}^n{road}_i}$
Enclosure	$Enclosure=\frac{{\sum }_{i=1}^n{building}_i+{\sum }_{i=1}^n{pole}_i+{\sum }_{i=1}^n{trunk}_i}{{\sum }_{i=1}^n{sidewalk}_i+{\sum }_{i=1}^n{fence}_i+{\sum }_{i=1}^n{road}_i}$

Note: The n refers to the number of street view images associated with each URLR. The i refers to the i-th street view image on the association. ${sum}_i$ is the number of all pixels in the ith street view image. ${vegetation}_i$ is the number of vegetation category pixels in the i-th street image. Other indications are analogous.

. Percentages of each component in the four-directional (north, south, east, west) dataset of each location covering various street views were summarized. The mean value for each location was calculated to represent the average condition of that URLR.

3.4. Image and Text Feature Fusion and Model Construction

3.4.1. Determination of EA Feature Weights

Different features have different degrees of overall impact on EA. The AHP [68] is combined with EEM [69] to identify the mean value for attaching weights. The EEM determines feature weights based on the magnitude of information provided by the quantitative value of each feature. The method reduces the subjectivity of weights and corrects the importance of objective indicators under small samples.

3.4.2. Feature Modelling of EA

Multimodal fusion is carried out with a single URLR, and the EA features of the URLR are extracted from the comment text and street view image corresponding to a single URLR, so the following representation model is constructed for multimodal feature fusion.

Each URLR will involve comments. The raw long comment data are denoted by T, and n is the number of long words involved in the URLR:

$$T=\{T_0,T_1,T_2,\cdots ,T_n\}.$$

(3)

To measure EAs at a fine-grained level, the long text is split into short texts, which are denoted by ST:

$$T_0=\{S{T}_{00},S{T}_{01},S{T}_{02},\cdots ,{ST}_{0n}\}.$$

(4)

Each URLR may be associated with street images. These street images are denoted by I, and n is the number of related images:

$$I=\{I_0,I_1,I_2,\cdots ,I_n\}.$$

(5)

The EA features of the URLR extracted from the text are denoted as $\{{Ft}_0,{Ft}_1,{Ft}_2,\cdots ,{Ft}_n\}$, where each component denotes the mean value of each EA feature dimension sentiment, and n is the number of subjective features. Similarly, the EA features of the URLR extracted from the image are denoted as $\{{Fi}_0,{Fi}_1,{Fi}_2,\cdots ,{Fi}_m\}$, where each component represents the quantized mean value of each EA feature dimension in all the associated street view images. Therefore, the final sequence of the URLR’s EA features is obtained as Formula (6):

$$\{{Ft}_0,{Ft}_1,{Ft}_2,\cdots ,{Ft}_n,{Fi}_0,{Fi}_1,{Fi}_2,\cdots ,{Fi}_m\}.$$

(6)

Each feature has a varying degree of influence on the overall EAs of the URLR, and the magnitude of the influence of each feature, i.e., the weight, is expressed as Formula (7):

$$W=\{w_0,w_1,w_2,\cdots ,w_{n+m+1}\}.$$

(7)

Finally, the quantitative results of the EAs of the URLR are calculated as Formula (8):

$$Ea={\sum }_{i=0}^n\left(w_i\times {Ft}_i\right)+{\sum }_{j=0}^m\left(w_{n+j+1}\times {Fi}_j\right)$$

(8)

4. Results

4.1. Environmental Amenity Features and Their Weights

4.1.1. Results of Feature Factors Identification for Text

We utilized 54,544 lengthy reviews as samples for extracting EA features. The segmented samples were used as a corpus, and a Word2Vec model was trained. We selected the top 100 high-frequency words and their semantically similar words, a total of 185 words for clustering. The distribution of words after dimensionality reduction to two dimensions and the cluster results are shown in Figure 9.

Words that are semantically similar in characterizing environmental features are clustered together. Still, there are a few misclassification cases because the Word2Vec model yields Chinese word vector representations containing contextual semantic relations. For more ideal clustering results, we sorted the clustering results and initially get the feature categories as sub-features based on the distribution characteristics of feature words. We further summarized the sub-features to obtain the main features, which are the category criteria for subsequent quantitative calculations. The EA features of the URLR including 15 sub-features and 10 main features were finally concluded, as shown in

Table 4. Pre-classification of environmental amenity features.

Feature Words	Sub-Features	Main Features
not far, distant, transportation, conveniently, parking, car parks…	Convenience	Convenience
lofty, spacious, bright…	Spaciousness	Physical setting
facilities, equipment…	facility
decoration, arrangement, design, layout…	Decoration
hygiene, tidy, clean…	Hygiene condition	Hygiene condition
style, characteristic, grace, architectural style…	Characteristic	Characteristic
one-stop, eat and drink, ‘eat, drink and be merry’, have all one needs, everything is there…	Functionality	Functionality
signboards, counters, old brand…	Grade	Reputation
too prosperous, prosperous, waiting for a seat, queue up…	Popularity
pedestrian flow, overcrowded, traffic…	Crowding	Crowding
kids, little ones, childhood…	Child-friendly	Child-friendly
expensive, price, cheap, value for money…	Price level	Consumption level
special offers, discount, group buying…	Strength of offer
taste, good to drink, menu, fresh…	Dining quality	Dining quality
service attitude, service, hospitality…	Service	Service

.

To obtain the classification confusion matrix, sub-features were selected as classification categories to eliminate the correlation among features. About 1000 samples were used for each class of EA features, and the final test accuracy reached 89.2%. The test set classification confusion matrix results are shown in Figure 10. The confusion matrix results show that the popularity and the price level have high misclassification rates with the crowding and the strength of the offer, respectively. We adjusted the feature classification for the former by moving the crowding into the popularity. The clustering diagram also shows that crowding and popularity are close to each other, indicating that the adjustment is reasonable. On the other hand, the more people a URLR has, the more popular it is. For the latter, the price level and strength of offer are part of the main feature of the consumption level and have no impact on the final measure, so no adjustment was made.

4.1.2. Weights Identification

In this study, with the help of the SPSSPRO platform, 30 experts scored and passed the consistency test with a consistency ratio of less than 0.1 for AHP. In addition, the weights for EEM were obtained by calculating the quantitative values of the features from a set of sample data. The results of the weight assignment are shown in

Table 5. Perceptual dimensions and their features and weights.

Sources of Perception	Perspective Weight (%)	Features	Weight (%)
Comment text	75	Convenience	11.60
		Physical setting	10.02
		Hygiene condition	15.44
		Characteristic	4.08
		Functionality	5.79
		Reputation	14.45
		Child-friendly	8.75
		Consumption level	9.05
		Dining quality	5.35
		Service	15.47
Street view image	25	Greenery	10.46
		Openness	26.66
		Enclosure	23.47
		Walkability	40.40

.

From the results, the perceived weight of comment text is greater than that of street view image. The comment text is mainly biased towards the interior space of a URLR. Overall, the EAs of a URLR are strongly influenced by service, hygiene condition, reputation, walkability, and convenience (Figure 11).

4.2. Spatial Distribution of Environmental Amenity Levels

Taking 2019, the year before the outbreak, as an example, it was classified into five levels by natural breaks based on the results of the comprehensive quantitative analysis of the environmental perception, as shown in

Table 6. Level classification of EA.

Level	Statistics Interval
1	0~0.472105
2	0.472106~0.522661
3	0.522662~0.567656
4	0.567657~0.634740
5	0.634741~1

. The spatial distribution characteristics are shown in Figure 12, which reveals three findings.

• Levels of EAs are characterized by clustered distribution. From the quantitative results, dark umber patches indicate a high-value range of EA, followed by saddle brown patches, mainly concentrated in the center of the main city district. Conversely, yellow and orange patches are in the low-value range for EA levels, and these patches are mainly found in the fringes of the main city district and the fringes of the suburban districts (Figure 12a). The EAs of URLRs in the Nanjing urban area show a spatial pattern of decreasing from the core to the outer edges, in which the core area presents a high level of spatial aggregation.
• The four urban districts show an uneven distribution of different levels. Comparing the district differences, the average level of EAs in the main city district (0.5613), which has the largest percentage of higher levels, is higher than in the three suburban areas (Figure 12c), with the suburban districts in the order of Dongshan (0.5481), Jiangbei (0.5210), and Xianlin (0.4983). Moreover, the main city has the most densely populated URLR patches. In contrast, Xianlin has the least populated URLR patches (Figure 12a,c), and no higher grades of URLR patches were found in the Xianlin district.
• Overall, the EA value of Nanjing’s URLRs is at a medium-high level (the average value of Nanjing’s URLRs is 0.5465), and the patches above the intermediate level account for 48.9%. From the point of view of the EA level distribution, more than three-fifths of the patches with medium and high EA levels (35% and 33%, respectively) show a pattern of low percentages in the two poles and a high rate in the middle (Figure 12b).

Further, Optimized Hot Spot Analysis by ArcGIS 10.6 is performed to obtain a high-level hotspot clustering map through the EA measure values, and the hotspot area is located at the junction of Xuanwu, Gulou, and Qinhuai districts, where major shopping malls are gathered, as shown in Figure 13. Surprisingly, the center of the aggregate polygon formed by the hotspot area is Xinjiekou Subway Station. Known as the number one shopping district in China, Xinjiekou is the commercial and economic center of Nanjing, and its transportation is very convenient, with 24 exits from the subway station. Secondly, it has comprehensive services, bringing together various places for shopping, eating, drinking, and entertainment.

4.3. Spatial Similarity Analysis of Environmental Amenity Features

To analyze the spatial similarity of EA features, this study was carried out with the help of the Grouping Analysis tool in ArcGIS 10.6 for the above 14 EA features, and the most suitable four groups were selected for analysis after several experiments. The results show a spatial stratification of EA features between the center and the periphery in Figure 14. From the spatial distribution of similarity patches, EA features are similar in each district’s edge areas (Figure 14a). From the statistical results of feature grouping, class I has the greatest difference in the distribution of each district, and it is mainly the feature of the main city district (Figure 14b). With regard to the distribution of classes in each district, Xianlin district has the most similar EA features; class I of the main city shows obvious central clustering, while the features of Jiangbei’s URLRs are different (Figure 14c).

4.4. Spatial and Temporal Variations of Environmental Amenity

To make comparisons over the period from 2019 to 2022, the quantitative measurements for 2022, when the epidemic is gradually liberalized, are ranked using the same numerical breakpoints as in 2019, and the results are shown in Figure 15, which reveals three findings, as follows.

• The EAs of Nanjing’s URLRs have declined overall. Firstly, as reflected in the interval of quantitative values of EA, the value has changed from [0.409060,0.808437] to [0.351678,0.765754]. Secondly, the number of higher levels of the first three has declined, and the number of lower levels of the last two has increased. It can be further seen from Figure 15 that all districts are experiencing a reduction in EAs at higher levels. This could be because the epidemic has reduced the total number of customer visits, there has been a subsequent decrease in comments, and even some leisure venues have closed. However, some URLRs’ EAs at some neighborhoods’ entrances have improved (Figure 16). This is due to the gradual liberalization of the epidemic in 2022; safe recreational activities with short distances satisfy people’s unsealed desires.
• The spatial distribution of grades in 2022 is similar to that in 2019, with a decreasing trend of multiple cores radiating outwards. However, the phenomenon of high-grade clustering in the main urban areas is more concentrated compared to 2019.

5. Discussion

With the economic development and people’s pursuit of a high quality of life, measuring the urban leisure environment has become a much-needed research direction, especially for the EAs of URLRs, which significantly impacts a city’s economic development. Previous research has focused on evaluating urban amenities and measuring the EAs of tourism. These large-scale and large-scope measures have the advantage of full coverage of an urban area, but they merely concentrate on analyses from macro perspectives. To perceive the EAs of URLRs more granularly, this study adopts a perception framework based on big data and selects URLRs in Nanjing as research areas to reveal the EA features and their mechanisms.

An important part of this framework is fully exploring the features that characterize the EAs of URLRs. Fourteen types of EA features were perceived. Compared with the expert knowledge approach, this paper focuses on the human-centered perspective to explore the complex relationship between people’s perceptions and leisure space. It explains which aspects of the leisure space are perceived to be more important for humans from the environmental psychology perspective. It can be found that five features, including hygiene condition, service, reputation, walkability, and convenience, have the greatest impacts on the perception of EA. Hygiene conditions relate to human health, service meets people’s spiritual needs, reputation makes URLRs more attractive, and convenience increases people’s choice of URLRs. In addition, although the perceived weight of the street view image is much smaller than that of the text, the feature of walkability perceived from the street view image has a larger influence. This shows the importance of walking space for URLRs.

To further understand the mechanisms of EA levels in URLRs, the measurements and the spatial patterns of EAs in Nanjing’s URLRs were also analyzed. The results are summarized as follows.

• Firstly, from the perspective of spatial distribution, the EA level of URLRs presents a spatial pattern of generally decreasing from the center to the outside and shows characteristics of regional differentiation.
• Secondly, regarding feature similarity, features of EAs in Nanjing’s URLRs show a spatial pattern of similarity in the center and at the edges of each district, respectively.
• Thirdly, comparing analyses of 2019 and 2022 suggests that the epidemic reduced the overall EA level of URLRs in Nanjing, which confirms the previous effect of hygienic conditions on EA. However, there is an increase in EAs for URLRs in the vicinity of residential neighborhoods.
• It is worth noting that the spatial distribution of EA level is consistent with the urban vibrancy [70,71] in related studies. That is, it has an attenuation characteristic from the city center. The closer URLRs are to the city center, the more accessible they are to the surroundings and the higher their vibrancy. As the result of this study, the center of the region formed by the hotspot URLRs is Xinjiekou Subway Station. To our knowledge, central areas of a city are the focus of government attention, with substantial resources invested in infrastructure development and recreational services. On the other hand, residents tend to be wealthier in central areas [72], which attract a higher concentration of talent [73], and have higher expectations for quality of life and spiritual pursuits. These indirectly lead to differences in the spatial distribution of EAs of URLRs. This study can well explain the cause of spatial heterogeneity of leisure activities [50].

The above analysis can serve as a valuable reference for urban construction and leisure planning. Here are several suggestions:

• Firstly, urban planners and decision-makers should prioritize the aspects that are most important to people. This can be achieved practically by focusing on constructing walkways and other facilities or increasing the reputation of urban regional leisure resources through media promotion. For example, TikTok is an up-and-coming online publicity platform through which you can create featured online publicity.
• Secondly, the government should ensure a balanced development of urban areas. They can prioritize the construction of leisure infrastructure and introduce talent policies for non-central regions (such as giving more subsidies for talent) to promote the simultaneous growth of the economy in urban regions.
• Thirdly, it is crucial to leverage the unique features of each urban regional leisure resource. Local businesses and governments should clarify their characteristics and corresponding weaknesses, making the characteristics deeply impressive for people based on complementary advantages and disadvantages. For example, special cultural and creative neighborhoods are a good direction for development.
• Lastly, it is wise to establish comprehensive measures to respond to special public events. Adjusting policies promptly in the face of events and adapting to market changes can minimize adverse effects to the greatest extent possible. Building systematic monitoring and early warning systems and improving early warning capacity is the most effective.

6. Conclusions

The results of this study implicate that urban planning and policymaking should fully consider fairness in leisure environment quality from a human-centered perspective. The case of EAs of URLRs in this study emphasizes the diversity of leisure spaces and the combination of subjective and objective views based on the theory of EA. The main contribution of this study is threefold:

• An advanced framework is proposed to evaluate the urban leisure environment. In this study, we proposed an advanced framework for quantifying and characterizing the EAs of URLRs on a fine scale based on image and text fusion perception. This EA estimation method is developed explicitly for URLRs and is more comprehensive and effective than traditional methods. The proposed framework could be extended to measure the quality of other types of leisure space, such as sightseeing and culture.
• The understanding of human perceptions of EAs in a large-scale urban environment is enhanced automatically and efficiently using web comment text and street-level imagery. The subjective and objective perceptions can be combined based on the review text and the associated street images of a single URLR. Using natural language processing and computer vision technology, the EAs of URLRs can be characterized comprehensively. The feature factors of the EAs for the internal and external spaces of URLRs are extracted. The feature factors and influence weights affecting the overall EAs of URLRs are analyzed, enriching the EA system and the connotation of leisure EA. These perception methods help identify which features are more significant for URLRs, providing a clear path for environmental improvement.
• The time and space distribution of EAs in URLRs of a developing country city is analyzed, and its formation mechanism is preliminarily diagnosed. Compared with studies with larger coverage, such as urban amenities, this study focuses on a more fine-grained level, i.e., the humanistic leisure field. It reveals EA’s spatial distribution and spatial-temporal change characteristics in Nanjing’s URLRs. The measurement results of this study can offer understanding of the potential spatial heterogeneity patterns of urban leisure activities. The case study could provide decision-making support for urban planning and management of leisure areas.

Despite the merits of this study, there are still some problems to be solved. The three main limitations of this study are as follows: (1) Although the perception and measurement in this study are based on big data fusion, the time alignment in this fusion is a coarse-grained year alignment, and image and text fusion based on finer-grained temporal alignment will be further investigated in the future. (2) The evaluation of the leisure area in this paper is overall, but the actual leisure area also includes a variety of sub-categories, such as hotels, massage places, etc. The influence weights of different EA features for various leisure places require further refinement. (3) There are two cases of spatial relationship between the recreation area and the road: the road is outside the URLR, and the road is in the URLR. Different spatial relationships may lead to distinct impacts on EA. In future work, we intend to further explore the effects of the different spatial relationships between streets and URLRs on the overall EAs of URLRs, which is of great value in improving the research on the environmental measurement of recreational areas.