Sustainable Urban Green Blue Space (UGBS) and Public Participation: Integrating Multisensory Landscape Perception from Online Reviews

Abstract

The integration of multisensory-based public subjective perception into planning, management, and policymaking is of great significance for the sustainable development and protection of UGBS. Online reviews are a suitable data source for this issue, which includes information about public sentiment, perception of the physical environment, and sensory description. This study adopts the deep learning method to obtain effective information from online reviews and found that in 105 major sites of Tokyo (23 districts), the public overall perception level is not balanced. Rich multi-sense will promote the perception level, especially hearing and somatosensory senses that have a higher positive prediction effect than vision, and overall perception can start improving by optimizing these two senses. Even if only one adverse sense exists, it will seriously affect the perception level, such as bad smell and noise. Optimizing the physical environment by adding natural elements for different senses is conducive to overall perception. Sensory maps can help to quickly find areas that require improvement. This study provides a new method for rapid multisensory analysis and complementary public participation for specific situations, which helps to increase the well-being of UGBS and give play to its multi-functionality.

1. Introduction

1.1. UGBS: Multi-Functionality and Sustainable Development

Research on green infrastructure (GI) has become a focus of spatial planning and sustainable development in recent years [1,2,3]. The European Commission defines GI as a strategic planning network with high-quality natural and semi-natural areas (including green and blue spaces and other ecosystems) that aims to provide a wider range of ecosystem services [4,5,6]. The European Union has taken the lead in launching a 10-year plan for GI across countries, regions, and cities, aiming to build a GI network through collaboration, protect urban forests, integrate GI resources such as urban parks, rivers, and streams, improve the ecological network, and provide UGBS for residents [6,7,8,9]. As green and blue spaces are frequently mentioned together, and they are also the objects that planners need to pay attention to together, the term green–blue infrastructure (GBI) has been widely accepted in recent years. GBI extends the connotation of GI; GBI is a system composed of different green and blue areas with different natural degrees, which is more in line with the inherent characteristics of today’s cities, so this study adopts the definition of GBI and UGBS uniformly [10,11,12].

UGBS’s functions include helping cities cope with climate change, protecting biodiversity, assisting in effective stormwater management, and protecting and restoring urban ecology [13,14,15,16,17,18,19]. At the same time, it coordinates social and cultural benefits, including education, social relations, cultural heritage, and entertainment, enhances landscape aesthetics, promotes social equity, and improves social well-being and human health [4,6,20]. UGBS provides residents and visitors access to nature and places for leisure, relaxation, and social activities. They play an important role in creating sustainable, resilient, inclusive, and competitive urban areas that contribute to local economic development [4]. Constructing UGBS and developing GBI also have become important strategies for realizing sustainable urban development as an effective means to coordinate environmental, social, and economic development [21,22,23].

Multiple types of UGBSs support multiple functions. Urban expansion encroaches upon original natural environment. Artificial park systems, urban forests, or river courses after treatment and landscape transformation in the built environment can provide a multifunctional UGBS for residents, which is important for BGI in the city [24]. In addition to protecting urban ecology, citizens can perceive nature and recover psychologically by visiting or passing urban greenways and street green spaces while commuting [25,26,27]. UGBSs, such as parks in the city, also carry out social and cultural functions and are, therefore, very important functional spaces in the city [7,28,29]. The process of urban expansion often invades the surrounding natural environment in addition to urban parks and greenways, but people are gradually becoming aware of the damage and impact it causes on native vegetation and ecological landscape. To preserve and restore local ecology, many cities have established country parks, wetland parks, or nature reserves located in the suburbs or outer suburbs [30,31] that mainly protect ecological diversity and provide habitat for many plants and animals. Accessory green spaces, roof gardens, private gardens, and informal green spaces of buildings are also important supplements for GI. They provide potential habitats for plants and animals in the city, which helps support the diverse ecological resources in the city [32,33,34]. In general, urban parks, community parks, greenways, gardens, botanical gardens, country parks, wetlands, and nature reserves constitute the main BGI of the city and are the most important providers of UGBSs. Strengthening identification and management and enhancing versatility are conducive to promoting sustainable urban development [35,36,37].

1.2. Physical Environment, Multiple Senses, and Subjective Psychology

There is a certain relationship between the production of the human senses and the physical environment. People’s sensory organs, eyes, ears, nose, mouth, skin, etc., obtain different experiences through receiving stimulation from the physical environment, thus generating a subjective understanding of the physical environment and different psychological feelings [38,39,40,41].UGBS can help with residents’ well-being and promote their physical and mental health; the physical environment plays an important role, especially natural landscape elements such as vegetation and water, which can significantly improve residents’ satisfaction [42,43]. However, as an increasing number of people are concentrated in cities or metropolitan areas, urban landscapes are filled with artificial elements, such as construction and transportation; people have increasingly lower contact with nature in daily life, and work stress in urban life has also contributed to a high incidence of public health events, including some psychological disorders such as depression and anxiety [44,45,46]. A lack of long-term outdoor exercise also increases the risk of cardiovascular and cerebrovascular diseases, obesity, dementia, and some cancers. UGBS in cities helps alleviate and prevent these diseases [47]. The sensory experience brought by nature is conducive to stress reduction, psychological recovery, depression and anxiety relief, and multisensory function promotion in children with autism [48,49,50].

The verification of the ecological benefits of the UGBS is usually based on the evaluation of objective indicators, while research on psychological function is dominated by the measurement of subjective indicators and the statistical analysis of psychological scales [51,52,53]. Among the evaluation studies on the effects of stress reduction and recovery, visual studies accounted for the largest proportion. The vegetation and natural elements provided by the UGBS benefit residents and visitors who have lived in fast-paced urban areas. Their stress levels are reduced by watching UGBS compared to people watching urban buildings and other elements, and they are more likely to achieve psychological recovery [54,55,56]. Therefore, the overall satisfaction and perception evaluation of the public will be higher with a better degree of greening, and they will have a stronger sense of happiness. Therefore, vision is considered to be an important sense for people to connect with the physical environment.

UBGS is an integrated space of natural and artificial elements. The physical environment it contains is not only related to vision but also may lead to a variety of other sensory experiences. There is a relationship between the production of human senses and their physical environment, and it is of great significance to evaluate residents’ overall perceptions and satisfaction after using them in urban decision-making and planning systems [57,58]. Subjective and comprehensive reflections on these dimensions revealed the overall emotional impression of space, which is the result of multiple senses working together, but the role of other senses besides sight is often overlooked [58,59,60]. In the process of design, management, and promotion, the UGBS should enhance people’s multisensory experience when visiting and using, which will help improve their perception of the overall atmosphere, receive the transmitted information in the process of space experience, and form a memory of the city image and self-understanding of the experience [61,62]. It is more comprehensive to integrate the public’s overall perception based on multi-senses compared to an assessment based on a single sense [63].

1.3. Online Reviews Complement Public Opinions and Participation

Understanding the subjective perception of the public is of great significance for enhancing the attractiveness of UGBS and promoting sustainable development [64,65]. In terms of government decision-making, public participation is a complex and controversial process, especially because marginal groups have fewer opportunities to express their views, making the effectiveness of participation results limited [66,67]. Public opinion reflected in online reviews is an effective indicator of the interaction between local governments and citizens and can fully reflect the views and emotions of the public. Online platforms are a relatively open environment, and different social groups have relatively fair opportunities to speak [68,69,70]. Therefore, public opinion feedback through online reviews can serve as a supplementary form of public participation. These are real experiences and feedback of multiple senses generated by actual contact with the physical environment in daily life [71]. Consequently, online reviews have become valuable data sources. Large volumes of data can be acquired relatively quickly. However, because of its unstructured information characteristics, current research mainly focuses on geographic coordinates and target opinions of data, which are difficult to statistically carry out [72], and there is also little research on information mining of different senses at present.

With the development of deep learning (DL) technology, it is possible to convert unstructured language data into language that computers can recognize, which can further and quickly mine the information from online reviews and evaluate the emotions behind them. It has been widely used in the processing of text and picture information [73,74,75]. For example, Koblet et al. obtained first-person assessment information on landscape features using online texts [76], and Huai et al. obtained terms of environmental features using Word2vec and word embedding technology to compare the effects of environmental features on overall perception under different cultural backgrounds [77]. However, there has been no research to explore different sensory information in online reviews.

In view of the above research status and problems, this study innovatively used the latest DL technology to realize data mining in online reviews and conducted physical environment and sensory analyses to solve the following problems:

• How do different sensory and physical environments affect the overall perception level?
• Is there a correlation between the physical environment and multiple senses?
• Compare the subjective perceptions of major UGBS in Tokyo based on online reviews.
• Supplement public participation and opinions with the overall multisensory perception to improve sustainable management.

Solving the above problems will help with understanding the differences in public perception of UGBSs in different districts, quickly finding the good or bad senses in different sites, and enhancing or controlling them to promote sustainable management, increase the versatility of UGBS, and promote residents’ well-being. This fills the gap of data analysis based on online reviews in the past.

2. Materials and Methods

The workflow is illustrated in Figure 1. We chose Tokyo (23 districts) as the case study area in this study (Figure 2) to understand the public’s perception of the UGBS.

Data in online reviews mainly include pictures and text; therefore, this study adopts two DL methods—Fully Convolutional Networks (FCN) and Natural Language Processing (NLP)—to separately process picture and text information in online reviews. The study fully mined pictures that mainly contained visual sensory information and text that may have different physical environments and multisensory descriptions. The sentiment score of the online reviews was rated using Google’s language platform, Cloud Natural Language (CNL). The results obtained were divided into different levels, which were used as proxies for the public’s overall perception. Finally, statistical analysis was used to study how different senses and physical environments affect public perception and evaluation and whether there is a certain correlation between senses and the physical environment. Sensory maps were created to provide a targeted optimization strategy for different sites.

2.1. Study Area

By the late 1980s, Tokyo had become a huge international city, and in the early planning of urban construction, a plan for the construction of an infrastructure system oriented toward the development of the manufacturing industry was determined [78]. Tokyo established a complete road transportation network and commercial hub, which is an inevitable demand for a period of high-speed economic development and rapid population growth [79]. Simultaneously, due to the highly dense urban buildings, most of the offices or residences are surrounded by an urban “gray landscape.” The dense layout of trams and other transportation networks makes them susceptible to noise [80]. Tokyo residents have a high proportion of commuters who often experience stress or depression owing to their work and relationships [80,81]. Tokyo’s high population density, intensive land use, and narrow roads have not left much space for road greening. Therefore, people mainly access public green resources and enjoy their social functional attributes or health benefits by visiting urban parks and other areas with relatively concentrated UGBS, which constitute the main BGI of the Tokyo skeleton [82].

2.2. Data Acquisition

Google Maps allows users to retrieve basic information and route guidance by searching for a destination. It also supports users with multilingual backgrounds in uploading reviews by a diverse population. Although there are some dedicated travel websites or social media platforms, such as TripAdvisor, Twitter, Facebook, Weibo, etc., they are usually considered valuable sources for obtaining the public’s real-time opinions on an event [77]. However, because they are tourism platforms or have untargeted social properties, some small UGBSs are not regarded as major tourist destinations. Therefore, such spaces may not receive attention from corresponding tourism review websites, resulting in less relevant evaluation. Google Maps is a mapping service that can search for any location, and visitors can rate a destination based on whether it is a scenic spot or only a daily and local UGBS. Other academic studies have demonstrated the effectiveness of acquiring public opinion using Google Maps [77,83,84,85].

Next, we screened the major UGBSs in Tokyo’s 23 districts on Google Maps. A total of 105 sites with at least 40 online reviews were selected, and the corresponding destinations were selected for each district. All procedures followed by the Google platform permit noncommercial academic research and do not require users’ personal information and disclosure (https://about.google/brand-resource-center/, accessed on 30 October 2022). Python 3.7.4 was used to batch-capture the text and pictures posted by users. After removing empty and repeated reviews, 20,236 pieces of reviews (including 70,000 pictures) remained.

2.3. Data Processing

Online reviews mainly consist of two forms of data: text and picture reviews. Traditional research often employs tourists to capture photos in order to acquire picture data; however, its efficiency is not high. Online reviews make it more convenient to obtain many pictures with tourists’ subjective preferences for analysis, saving money and time. Pictures taken by tourists mainly reflect visual emotions, the visual sensory preferences of tourists, and their physical environment.

Textual reviews include richer emotional feedback, as well as descriptions of the physical environment and sensory experience. However, text data are unstructured and often cannot be recognized by computers or used for direct calculations and analyses. Therefore, effectively extracting and transforming picture and text information into computer-recognizable data is the key to this study.

In view of the above-mentioned data characteristics, the following three methods will be adopted in this study to process the data:

2.3.1. Sentiment Analysis

Currently, NLP is widely used in many fields, such as machine translation, spam detection, and medical information extraction [45,86,87]. However, little research has been conducted on public opinions on the UGBS. Sentiment analysis, also known as subjective surveys or sentiment artificial intelligence, is a kind of NLP that deals with the form of text through which information patterns and key features can be extracted from many texts to judge an author’s views and opinions. It can also be used to identify and analyze any data in text form, from product reviews to short posts on social media or review articles. CNL (https://cloud.google.com/natural-language/docs/analyzing-sentiment, accessed on 20 March 2023) is Google’s intelligent analysis platform that can conduct sentiment analysis of texts provided by users to determine the dominant views in texts, especially whether the author’s attitude is positive, negative, or neutral. After analysis by CNL, a sentiment score is obtained to represent the author’s overall perception, which is generally between −0.1 (negative) and 1.0 (positive); after normalization treatment, the score is between 0 and 1, then, using the natural breakpoint method, the sentiment score is divided into five levels. Python was used to make application programming interface (API) calls to the CNL and obtain the sentiment score of each review. Since Google Maps is an open service platform for global users, there are authors with multi-language backgrounds. We used Google Automatic Translate to conduct language unification. People’s overall perception of UGBS can be measured by sentiment analysis, and fuzzy text expressions can be quantified by a score, which we used as a proxy for overall perception.

2.3.2. Text Processing

The sentiment analysis provided by CNL is a mature application based on NLP. Users can independently conduct a sentiment survey of the text based on the platform algorithm. Public descriptions of the physical environment and senses contained in text data are of great significance in understanding subjective perceptions after accessing UGBS. However, there is no mature application that can directly complete such a targeted assessment. Therefore, we need to combine NLP and principal component analysis to screen the target words. Google recently introduced Bidirectional Encoder Representations from Transformers (BERT), which is an innovative technology that has been pre-trained on large lexical datasets and fine-tuned with an additional output layer based on traditional DL [88,89]; thus, it can output new models for various NLP tasks [88,90], as shown in Figure 3. Bello et al. used a pre-trained transformer, BERT, together with a convolutional neural network (CNN) and recursive neural network (RNN) to verify its state-of-the-art performance [90].

For this study (Figure 1), we first removed duplicates and empty revisions at the data acquisition stage. Then, we removed useless and inaccurate data, including spelling errors and non-target language. Jieba was used for tagging, stem extraction, and part-of-speech recovery, followed by the removal of the stop word using the reference stop thesaurus. Through this process, we gained a wealth of tokens. We used BERT to translate them into a target language that computers could recognize, which contains large eigenvectors. However, the word vector obtained by this process is a feature vector with hundreds of dimensions. To find a good physical environment and sensory description that will affect visitors’ subjective perception, in combination with the purpose of this study, we needed to use manual screening. Principal component analysis (PCA) is a data processing method based on principal component eigenvalues. This method maps the multidimensional data into low-dimensional space and characterizes the comprehensive characteristics of the research object in the form of several principal component factors. The advantage of this method is that it can reorganize and extract the characteristic information of all research objects to obtain the most representative principal component factors instead of filtering the original variables [92]. Therefore, we used PCA to reduce the dimensions and project the coordinates of the word vector after dimensionality reduction onto the two-dimensional coordinate system. K-means clustering algorithm (K-means) is a kind of unsupervised clustering analysis algorithm using the iterative solution, which is widely used. By using K-means, we screened out the main target clustering [93]. In this study, python’s algorithm toolkit Scikit-learn was used to complete these two steps.

According to previous studies on landscape feature evaluations, the physical environment of a UGBS can be divided into two categories: natural and artificial [77,94,95]. Then, combined with clustering, the physical environment was also subdivided into these two classifications. For each category, keywords with the top 5–7-word frequency were selected. These core words were used for lexical extension, and words (vector cosine less than 0.5) were selected next [77,96,97]. Similarly, word clusters related to sensory description were found in the clustering vocabulary, core keywords were screened, and a list of feature words was created. However, it is worth noting that words that reflect the physical environment are mostly nouns, whereas clusters that reflect the senses are mostly adjectives. However, to obtain a comprehensive description of the senses, organs that can receive sensory signals and objects that elicit sensory reactions were also taken into consideration.

2.3.3. Picture Processing

Picture intelligence recognition technology and language processing are the basis of computer calculation, processing, and feedback. Semantic segmentation is an important technique in this field. It can transform pictures into languages that can be recognized and processed by computers. Intelligent recognition and semantic picture segmentation technologies are widely used in landscape planning [98]. For example, researchers have used Google Street View maps to calculate the visual greening rate from a human perspective after semantic segmentation [99]. With the help of DL, they can quickly identify and calculate the average visual greening rate of an entire street, which can help analyze the influence of vegetation on various perception evaluations. In this study, after FCN processing, the pixel ratio of each landscape element was calculated; for example, the vegetation pixel ratio (VPR) = Vegetation pixel/whole pixel (other landscape elements were similar) as a representative of the greening rate of human perspective. The results reflected the visual perception of the public and the actual physical environment [100]. Many years of data can also be obtained for analysis, which greatly expands the amount of data, reduces the difficulty of data collection, and improves efficiency much more than manual collection and analysis [98,101]. However, the pictures in online reviews are different from those in the Street View map, which are captured at a constant speed and orderly intervals by moving vehicles bound to cameras with relatively fixed heights and angles. Pictures in online reviews are open and unrestricted. Different authors have used different equipment and shot at different angles, which causes the form, quality, and size of the pictures to vary, making picture recognition difficult. Based on the above problems, we chose FCN as the picture processing tool, as it can realize pixel-level recognition and is currently one of the most selected models in semantic segmentation and picture recognition tasks [102]. The difference between an FCN and traditional CNN is that the last layer of an FCN is not a fully connected layer that enables the model to integrate information. Instead, the final number of output channels was modified using convolutional networks. The main advantage of this method is that the model is not restricted by the full connection layer, so the size and form of the picture we can input are more flexible [98,102] (Figure 4). The ADE_20K dataset is annotated and published by MIT CSAIL Computer Vision Group (http://groups.csail.mit.edu/vision/datasets/ADE20K/ accessed on 20 March 2023). The dataset contains 150 object categories based on annotations of 25k multi-scene photos, which can be used to analyze the complex and multi-scene photos generated by visitors; it can cover the categories of landscape elements in previous studies, which is very authoritative and general [103,104]. The ADE_ 20 K dataset was used for training and validation, and the types of elements were merged or deleted according to the purpose of this study and previous studies (Appendix A). For the elements that appear in the database and can be identified but are not the main research objects in the previous literature, we summarized them into “others.”

2.4. Sensory Map

The geographical locations of UGBSs in Tokyo’s 23 districts differ—some are in bustling urban areas, adjacent to the main traffic road, and surrounded by an office environment; some are in a community with a relatively small service range; some are in the sea; and some are located inland. Sensory maps for different geographical locations are helpful for managers and planners to intuitively understand and improve according to different environments. Finally, suggestions for improvements based on geographical and environmental factors are provided.

2.5. Statistical Analysis

Pictures and text processed by DL were transformed into computable feature data, including physical environment features and sensory descriptions. Next, statistical analysis can be used to solve the problems mentioned in the introduction.

2.5.1. Correlation Analysis

The Pearson correlation coefficient was used to evaluate whether there was a correlation between the public’s perception of the physical environment of UGBS and sensory expression, and the significance level was set at p < 0.05.

2.5.2. Multiple Linear Regression Analysis

The public’s overall perception level acted as a proxy for the sentiment score. We needed to explore how different physical environments and senses affect overall perception. Therefore, multiple linear regression analysis was used to take the overall perception as the dependent variable and the physical elements of secondary classification in the physical environment and sensory description as independent variables to determine how they contribute to overall rating satisfaction or perception levels.

3. Results

3.1. Public Overall Perception

In Figure 5, the twenty-three districts (a) and each UGBS site (b) are divided into five levels. There are differences in perception between districts and sites, but the public’s average sentiment score (SSAVG) for the selected sites was mostly positive; the standard deviation of the sentiment (SSSD) score shows a difference (c). Taito (SSAVG = 0.906, SSSD = 0.01) had the highest average score, followed by Bunkyo (SSAVG = 0.903, SSSD = 0.02), Nakano (SSAVG = 0.896, SSSD = 0.03), Edogawa (SSAVG = 0.810, SSSD = 0), and Ota (SSAVG = 0.797, SSSD = 0.1). For a particular site, Sarue-onshi Park (SS = 0.945) and Furukawa Garden (SS = 0.943) had the highest emotional scores. Okubo Park (SS = 0.681) and Keihinjima Tsubasa Park (SS = 0.571) scored the lowest.

3.2. Text Data

The information contained in textual data was divided into three categories: natural elements, artificial elements (including landscape elements and functional services), and senses, as shown in Figure 6 and Figure 7.

Of the middle-upstream levels (SSAVG > 0.8785), Kita (SSAVG = 0.883, SSSD = 0.035) had the most descriptions of natural landscape elements in the text, mainly “flowers” and “water,” but also some “animals” and “mountains.” Second, there were also a lot of descriptions of “functions” and “people.” Combined with the number of reviews and sentiment scores, Furukawa Garden (SS = 0.943), as the representative site, creates a better natural environment, which is very attractive, and may improve the average level. The high perception level of Bunkyo (SSAVG = 0.903, SSSD = 0.02) may also be related to natural landscape elements since the three sensory experiences of visual, auditory, and taste are more balanced. In contrast, Kita’s multisensory experience was relatively weaker than Bunkyo’s.

Of the intermediate levels (0.8625 < SSAVG < 0.8784), Chuo (SSAVG = 0.874, SSSD = 0.032) was the most prominent in gustation. This may be because Hamachō Park (SS = 0.842) is an attractive site for picnics and vegetation, whereas the other senses are dull. The same was true for Adachi (SSAVG = 0.878, SSSD = 0.001). Arakawa (SSAVG = 0.087, SSSD = 0.002) provided more descriptions of vision and hearing, but other senses were relatively weak.

In the middle-downstream area (SSAVG < 0.08624), the visual description of Toshima (SSAVG = 0.0862, SSSD = 0.017) was particularly outstanding and provided certain auditory and taste experiences with rich senses. However, the overall perception level was in the middle-downstream area, which may be related to the fact that the built environment was too dense and the natural environment was scarce. According to the results of the picture and the corresponding text, Ota (SS = 0.797, SSSD = 0.1), the physical environment of this area is mainly empty, with little vegetation and architectural elements. Jonanjima Seaside Park (SS = 0.755), Keihinjima Tsubasa Park (SS = 0.571), and Nishi-Rokugō Park (SS = 0.789) have relatively poor sentiment scores.

3.3. Picture Data

As shown in Figure 8, vegetation, sky, and water were the main natural elements photographed, which is consistent with the function of the UGBS. Animal elements are less expressed in the picture because it is difficult to see them in ordinary UGBS. Among artificial elements, buildings, roads, and certain interior elements are easily perceived and recognized by visitors. However, few landscape structures, such as cultural sculptures or landscape sketches, have been recognized, which may not be very attractive to visitors or not enough in most UGBSs.

The expression of each landscape element differed among the 23 districts in Tokyo. The vegetation in the Bunkyo (VPR = 0.466) and Minato (VPR = 0.451) districts was reflected more in the subjective cognition of residents, indicating that there is more vegetation in the objective environment of these two districts, which is also favored by tourists. Edogawa (WPR = 0.122) had a higher proportion of water elements, indicating that the water bodies in this area were mapped more in tourists’ subjective cognition. Toshima (BPR = 0.241) and Sumida (BPR = 0.177) had higher proportions of architectural elements and less vegetation. Shibuya (PPR = 0.048) and Bunkyo’s (PPR = 0.042) UGBSs were more densely populated, as they are among the most densely populated areas in Tokyo. Ota (SPR = 0.330) is far away from the city center and has less natural vegetation and architecture; it is mainly open space. Various elements, including natural and artificial elements, show differences in different districts and UGBSs.

3.4. Sensory Map

The sensory results of the 105 sites were plotted as a map (Figure 9). Because there are very few descriptions of olfaction, it was combined with gustation.

Furukawa Garden (SS = 0.943), Ogunohara Park (SS = 0.891), Zempukuji River Green Space (SS = 0.878), Odaiba Marine Park (SS = 0.886), and Tokyo Metropolitan Hikarigaoka Park (SS = 0.925) had higher sentiment scores. Combined with the results of the physical environment analysis from the pictures and text, these sites had good greenery, flowers, water features, and other natural landscape elements.

Furukawa Garden (SS = 0.943) and Ogunohara Park (SS = 0.891) had the highest auditory expression and sentiment scores. Natural sounds such as running water, birds, and insects play an important role in visitors’ relaxation and pro-nature experiences. Heiwanomori Park (SS = 0.866) mainly provides places for sports (such as running tracks) and a natural environment for residents. Although visitors’ visual perception of the site is not very prominent because the site is open and quiet, they can hear many birdcalls and get a sense of a “quiet” experience. Tatsuminomori Seaside Park (SS = 0.763) and Yumenoshima Park (SS = 0.775) have been under construction and maintenance for a long time, resulting in noise that affects the overall perception level of visitors. Jonanjima Seaside Park (SS = 0.755) is susceptible to aircraft noise, owing to its proximity to the airport.

In terms of olfaction and gustation, the roses planted in Furukawa Garden (SS = 0.943) give a higher perception. The site provides tea rooms and special ice cream, which allows visitors to obtain a rich sense of smell and taste. Hamarikyu Gardens (SS = 0.906) have created an imperial teahouse in the middle of the central water feature, where visitors report that they can be physically and mentally restored by tasting distinctive matcha and refreshments and admiring the water features of the Japanese courtyard. The Itabashi Botanical Garden (SS = 0.811) offers visitors a café where plants can be seen in a greenhouse, which many visitors recommended as unique. Odaiba Marine Park (SS = 0.894) is a coastal site and a typical blue space where visitors can obtain a visual experience of the waterscape, hear the sea, and experience the physical comfort brought about by the sea breeze. However, some authors mentioned that they could not withstand the taste of seawater; therefore, they gave low evaluations.

In terms of feelings, Odaiba Marine Park (SS = 0.894) is a coastal landscape; sea breeze and sunshine provide visitors with a good physical feeling, and the tactile feeling of stepping on the beach can also relax them. The Itabashi Botanical Garden (SS = 0.811) is hot and stuffy because it is located in a greenhouse. On the other hand, Nishi-Rokugō Park’s children‘s rubber ornaments are exposed to the sun, turning them hot after exposure, causing some adverse perception feedback.

3.5. Physical Environment, the Sensory Expression, and Overall Perception

3.5.1. Correlation Analysis

According to the text results of this study, Figure 10a, the physical environment may cause different sensory experiences in visitors, affecting their overall perception level. To determine whether there was a correlation between the physical environments and overall sensory perception, we calculated the Pearson coefficient and drew a heat map, indicating that there was generally a significant correlation between the sensory and physical environments (p < 0.05). However, the Pearson correlation coefficient between a single sense and the physical environment was not high (r < 0.3), indicating that there may not be a brief linear correlation but complex nonlinear relation.

Among them, vision has a positive correlation with both natural and artificial physical entities but is more strongly related to greening; hearing has a positive correlation with water, animals, people, activities, etc.; gustation has a positive correlation with the artificial environment; and olfaction has a positive correlation with greening and flowers. These landscape elements can be appropriately considered in environmental optimization to enhance the sensory experience.

As for picture data, Figure 10b, there were correlations between different landscape elements, among which there was a negative correlation between three visual elements: architecture, sky, and vegetation. The elements of mountain and water have a positive correlation, and tourists will be more inclined to the combination of these two elements. Among artificial elements, architectural and interior elements are often present, and human and artificial elements have positive correlations. Tourists tend to photograph scenes rich in vegetation and landscapes that match mountains and rivers, followed by architectural elements that are more closely related to human activities.

3.5.2. Multiple Linear Regression Analysis

First, we examined the normality of model residuals, ANOVA, and multicollinearity using the Kolmogorov–Smirnov (KS) test to solve the multicollinearity problem between the predictor variables. The test results show that the residuals follow a normal distribution, p > 0.05. The occurrence of a model tolerance of <0.2, or a variance inflation factor (VIF) > 10, is indicative of a potential multicollinearity problem, and the current VIF value is acceptable [105,106]. From Figure 11 and

Table 1. Summary of the regression models predicting perceptions in Tokyo (n = 105).

Data Source	Elements	Coef	Std Err	z	p > |t|	VIF
Pictures	Sky	−0.0175	0.02	0.65	0.037	2.056
	Vegetation	−0.0205	0.018	1.589	0.024	2.827
	Mountain	0.0288	0.047	−0.679	0.054	1.169
	Water	−0.01	0.03	2.063	0.074	1.395
	Animal	0.6121	0.604	0.007	0.031	1.006
	Architecture	−0.0127	0.024	0.729	0.06	1.811
	Structures	−0.0281	0.059	0.347	0.064	1.128
	Road	−0.0341	0.029	2.224	0.025	1.335
	Transportation	−0.1882	0.145	0.244	0.019	1.026
	People	−0.0765	0.055	1.627	0.017	1.225
	Food	−0.2187	0.339	0.012	0.052	1.026
	Interior	0.0533	0.043	0.021	0.022	1.367
	R2 = 0.205
Multi-sensory description	Vision	0.0229	0.006	3.817	0	1.172
	Hearing	0.0707	0.007	10.1	0	1.066
	Gustation	0.006	0.004	1.5	0.122	1.172
	Feeling	0.0824	0.007	11.771	0	1.134
	Olfactory	−0.067	0.038	−1.763	0.08	1.038
	R2 = 0.281
Text-Natural elements	Greening	0.0195	0.003	6.5	0	1.296
	Flower	0.0503	0.003	16.767	0	1.215
	Mountain	0.0236	0.009	2.622	0.006	1.083
	Stone Sand	0.0707	0.007	10.1	0	1.066
	Water	0.0124	0.003	4.133	0	1.263
	Season	0.0378	0.005	7.56	0	1.193
	Nature phenomenon	−0.0004	0.005	−0.08	0.028	1.152
	Weather	0.0198	0.011	1.8	0.063	1.059
	Animal	0.0118	0.004	2.95	0.001	1.192
	R2 = 0.226
Text-Artificial elements	Architecture	0.013	0.004	3.25	0.004	1.226
	Structure	0.0042	0.006	0.7	0.055	1.354
	Place	0.0354	0.006	5.9	0	1.221
	Road	−0.0134	0.007	−1.914	0.06	1.126
	Service	−0.0319	0.006	−5.317	0	1.201
	Transportation	0.0175	0.005	3.5	0.001	1.219
	People	0.0205	0.003	6.833	0	1.508
	Festival	0.0042	0.011	0.382	0.003	1.12
	Art–Culture	0.0895	0.014	6.393	0	1.037
	Activity	0.0619	0.016	3.869	0	1.04
	Expenses	0.0286	0.01	2.86	0.004	1.083
	Function	0.0645	0.003	21.5	0	1.625
	Sports	0.0009	0.008	0.112	0.009	1.125
	R2 = 0.235

, the picture information provides a prediction effect of 20.5% in Figure 11a, the sensory description provides a prediction effect of 22.6% in Figure 11b, and the description of natural elements in Figure 11c and artificial elements in Figure 11d in the text explains 22.6% and 23.5% of the model, respectively. Except for natural phenomena, all other natural factors play a positive role in prediction. The positive prediction effect of flowers was the strongest, and there was little difference in the feedback of each site, indicating that visitors of different sites would have a positive overall perception because of flowers, while the difference and perception of greening were relatively significant, indicating positive predictors of greening that were not high in some site.

Service plays the most negative role in predicting perceptions, followed by roads among artificial factors. Visitors pay more attention to these items, including parking, public toilet, etc. Meanwhile, cultural and art activities, exhibitions, etc., can meet the cultural needs of residents and play the most positive forecasting role, which also reflects the residents’ demands for the multifunctional space.

In the sensory description, smell plays a negative role in predicting perceptions, while others play a positive role. The auditory and somatosensory senses exceed the visual sense, indicating that they should be given more attention, although they were often overlooked before. Among them, the description of smells in different sites differs greatly, so individual sites need to avoid bad smells.

4. Discussion

4.1. Public Subjective Perception and Sustainable Development of UGBS

In this study, it is found that some sites show an imbalance of UGBS overall perception (sentiment score), which is not conducive to green justice. At the same time, the existence of certain unfavorable senses will also significantly affect the overall perception level, and a good sensory experience as diverse as possible will promote the overall perception, which can be achieved by optimizing the physical environment, especially the role of natural elements, and controlling unfavorable artificial elements (such as odor, noise, etc.) and appropriately increasing favorable artificial elements (such as cultural facilities, activities, etc.). The role of flowers in different sites is basically its positive prediction effect, but greening may not be absolute. There is a need for better artificial supporting facilities and service experience, so it is necessary to improve both natural and artificial environments and improve the management level.

Currently, there are several cases in Europe and other countries regarding public participation or integrating overall public perception and evaluation into local decision-making, which is conducive to the sustainability of UGBS planning [107]. Through an analysis of 23 districts in Tokyo, we can observe differences in subjective feedback among them, which can be regarded as a form of passively acquired subjective perception of the public, and to a certain extent, as supplementary to public opinion. Because the DL analysis method is adopted in this study, the expression of different senses can be quickly analyzed, which can effectively improve the efficiency of public participation.

The urban environment is a complex system. In the process of management of UGBS, not only a single sense of the public should be considered, but also the multi-sense experience, which will help expand good experiences to enhance the overall satisfaction and reduce the unfavorable sensory experience, which can further promote the well-being for residents provided by UGBS.

4.2. Sustainable Management Strategies Based on Multiple Senses

The results showed that the description of visual perception was the most common feedback information available from online reviews, and photos were mainly the embodiment of objective and subjective physical environments. People tend to photograph aesthetically pleasing natural objects, particularly greenery. The role of vision in this sense accounts for the most, which was also reflected in this study. Visual environment and experience make the greatest contribution to the overall perception. The perception of color occupies a very large proportion of the description, whereas shape and other visual elements are not obvious. Therefore, greening can be appropriately increased in urban spaces, especially with colorful flowering plants, to enhance the visual experience. However, vision was not the most significant contributor to positive evaluations, suggesting that there is still room for improvement in the overall visual experience. Especially for the areas with high building density, more greening should be provided to enhance the experience of natural elements, which will have a more obvious role in improving the overall perception.

Hearing also showed a positive effect, suggesting that the UGBSs in Tokyo’s 23 districts were conducive to providing a sound environment. Tokyo is a bustling city with highly developed rail transit and commerce. Coupled with a high population density and frequent activities, the sound of traffic and construction fills the city. High-decibel sounds increase pressure or make people irritable, while natural sounds, especially those of birds or running water, make people feel relaxed and even peaceful. Thus, people tend to visit UGBSs in cities to find inner peace and relaxation, which helps them escape from the city’s hustle and bustle for a while. However, some reviewers hope to have a “lively” experience, especially on holidays or when there are activities held, and they show a high level of cognition to the crowd flow and lively atmosphere brought by the activities held in the space. Therefore, artificial sound is not just “noise.” For example, artificial “music” can lead to a good experience. To control environmental sound, part of the noise from the source can be managed, or vegetation added to create an isolated space, which can play a good sound insulation effect to reduce the negative impact of noise. The sound that can bring a good experience to the UGBS, such as creating water features or facilities, can also enhance this. The experience of birdsongs can be obtained by increasing the habitat of birds or playing acceptable melodies in a suitable space. Some scholars have suggested that a good visual environment can help reduce the adverse effects of noise [108,109]. For example, Jonanjima Seaside Park (SS = 0.755) had some deficiencies in site management because of its proximity to the airport and the inconspicuous experience of other sensory environments. This results in low emotional scores. An enriched vegetation landscape and improved service and management of the site can compensate for the lack of an auditory sense of this site to a certain extent [40].

UGBSs generally carry out the daily leisure functions of residents, and the public’s perception of taste comes from refreshments provided by nearby businesses or parks. The formation of a gustatory experience in UGBSs is often related to services provided by the artificial physical environment. Taste plays a very important role in tourists’ perceptions of the city atmosphere, and special local food can promote their appreciation of this perception [110,111]. Simultaneously, we can also see that picnics in parks can also promote positive perception to a certain extent. The natural environment can increase visitors’ sense of happiness and pleasure when tasting food. Therefore, it is appropriate to provide residents with appropriate places for picnics, catering services with local food, or places to enjoy tea and rest that may provide a higher emotional value to visitors.

Physical sensory impression is mainly due to bodily sensations caused by temperature and humidity. Excessively high or low temperatures may cause different degrees of physical sensory impression, thus affecting overall evaluation. Touch can promote the development of healthy children and, to some extent, the recovery of children with autism. For example, touching a lawn, sand, or wooden stakes can promote the development of tactile senses in children, but these senses are not mentioned in current online reviews. This indicates that it is difficult for residents to acquire this sensory experience from the UGBS at the chosen sites or that it is a sensory experience that is easily ignored and, thus, not mentioned. However, vegetation provides a shading function in terms of temperature. The space created by such vegetation can bring residents or tourists a “cooling” experience, which can help alleviate the negative perceptions brought about by the urban heat island effect.

Some studies have proposed that the olfactory experience of fragrance can be obtained in the process of perceiving nature, which is mainly caused by natural elements such as vegetation and flowers [62]. These olfactory perceptions are difficult to obtain in the on-site environment. For example, fragrances emitted by forest vegetation are produced by natural chemicals. They are difficult to replace either in a simulated form or using artificial chemicals. However, many people with pollen allergies may not be able to obtain better olfactory perception at the site. Therefore, attention should be paid to avoiding high pollen counts when choosing plant materials. In other studies, visitors have been found to be very sensitive to olfactory perception. For example, public toilets and smelly rivers instantly trigger bad mood. General olfactory experiences seem difficult to detect, but a bad or irritating smell can easily trigger a negative evaluation. However, this study shows that the overall olfactory perception in Tokyo is a negative predictor. At present, several sites with low sentiment scores mention the odors of public toilets and rivers, indicating that some venues need to manage bad odors as soon as possible.

4.3. Optimize the Physical Environment Based on Sensory Feedback from the Site

With the help of the sensory map, it is possible to quickly obtain a prominent description of the sensory situation of each site, whether positive or unfavorable, which can help decision makers and managers to make integrated decisions.

The sensory experience and landscape elements of the UGBS physical environment that are subjectively accepted by the public in each district are different. The visual sense has some expression in each area, whereas smell exists only in a few areas. There were some differences in the descriptions of hearing, taste, and feelings. Regions with auditory descriptions include Bunkyo, Toshima, Suginami, Arakawa, and Ota, whereas other regions are rarely involved. The descriptions of taste sensations in Minato, Shinagawa, and Suginami are fewer than that in other areas, which may be because visitors in these areas have weak impressions of them. Notably, Kita, Nerima, and Adachi’s descriptions of somatosensation are limited, and visitors’ response to somatosensory stimuli, such as temperature, is not obvious, which may be because these areas are in the outer suburbs with rich vegetation or a certain blue space, and the heat island effect is weak. UGBSs with high sentiment scores generally have multisensory descriptions, and those with low scores may have 1–2 unfavorable senses.

In previous studies, residents who visited UGBSs for long periods reported higher satisfaction, happiness, and other physical and mental benefits [42,47,48,112,113]. As the UGBS provides the public with a multisensory experience of the natural environment, it is conducive to stress reduction and psychological recovery. Researchers later found that volunteers also showed high levels of perception and healthy physiological feedback when given a single visual stimulus, such as viewing VR or photos in the laboratory. This shows that the natural environment plays an important role in improving public perception, especially the visual aspect [97]. A team of researchers simulated hearing in the lab and found that natural sounds increase overall satisfaction, reduce volunteers’ stress levels, and lead to faster psychological recovery [114]. Although scents are not easy to collect, researchers have found that natural plant fragrances can bring feelings of physical and mental pleasure [62]. Other teams have shown that touching natural objects can promote a positive evaluation [63]. Overall, the sensory experience brought about by the natural physical environment was relatively positive. Artificial elements reflected an alienation, some of which played a positive role, while others had an obviously negative role. This study drew a similar conclusion, although different sites showed differences.

According to the results of the correlation analysis, all senses were correlated with the physical environment, but there may be some nonlinear correlations, which can be explored further. However, focusing on managing and creating a good physical environment or strengthening people’s sense of connection with such a physical environment through activities will undoubtedly promote positive sensory experiences. In contrast, in the city, bad sensory experiences lead to negative feedback or bad sensory emotions, which should be considered in the process of UGBS management.

In previous studies, a large proportion of architectural elements tended to be detrimental to the psychological recovery of visitors or reduce their overall perception level [115,116]. This study found that this is not absolute. A well-built environment, rich natural landscape, and good management services are conducive to a high level of perception. Some people do not think that buildings outside the city will affect their perception and overall perception of a certain UBGS and may even think that it is the embodiment of local culture.

The contrast between noise and quiet gives a strong sense of contrast. Because of their convenient locations, Bunkyo, Taito, and Chiyoda can quickly relax their hearing and obtain fresh natural air, which promotes a good experience for all senses. However, if artificial elements such as buildings in a UBGS account for too much, then there are almost no natural elements in people’s visual environments, which is detrimental to improving their overall perception level. In this case, people will pay more attention to the function, management, and service of the site, such as Toshima and Shinjuku, and be more sensitive to other senses in the environment, such as noise, bad smells, and temperature. It is suggested to appropriately increase vertical or small green space and control urban noise, which will help improve the overall perception. Their higher visual greening levels may be why Minato and Shibuya had slightly higher perception levels. However, although areas far away from the center have a better natural environment, rich vegetation resources, and water bodies that are conducive to the overall perception in traditional studies, they do not promote the overall perception level. Instead, visitors will put forward higher requirements for ecology and management, such as river water quality, environmental noise, odor, and the tactile touch of facilities. At the same time, we found that architectural elements are not a total disadvantage in surrounding areas such as Nerima, Kita, and Setagaya, which have a higher perception level than Ota, Katsushika, and Koto. The distribution and preference of residents for natural and artificial elements in the UBGS are not absolute in the overall perception. Thus, it needs to be based on the specific site situation and the actual needs of the surrounding residents for analysis. Rapid handling of public opinion through DL is conducive to swiftly analyzing specific environments and problems and promoting the sustainable development of the site.

5. Conclusions

In integrating UGBS planning into sustainable urban development, the main framework is the construction of BGI, including urban parks, urban forests, greenways, and riverfront landscapes. In the process of artificial construction of this “nature,” the public’s well-being and subjective feelings should be preliminary. If only a single sense is considered, some important aspects of management may be missed. Online reviews are easy to access and have large amounts of data. With the help of NLP and FCN, two types of data can be processed in large quantities, and useful information can be mined from them for multisensory and sentiment analyses. UGBS plays an important role in promoting sustainable urban development. On the basis of previous research on analyzing public opinions, this study distinguished different senses, which can help to solve existing problems in a targeted way.

The public’s subjective perception of UGBSs in Tokyo’s 23 districts is quite different, indicating a certain green inequity. For example, Bunkyo’s cultural heritage green–blue space will receive more attention and funds, thus completing better conservation and management, driving economic development, and bringing visitors more sensory experience and cultural feelings. Toshima has a large population, but there is very little space for UGBS, but in fact, people here need more green restoration and quiet natural areas. In remote areas such as Edogawa, although the UGBS is large, the level of management needs to be improved.

Rich sensory experiences often promote a higher level of perception, especially when brought about by natural elements. From the perspective of promoting the overall perception level, in the current environment, vision is no longer the most important factor, and the contribution of other senses plays a greater role, which should be paid more attention to promoting green equity and sustainable development, especially hearing and somatosensory senses. Even if only one adverse sense exists, it will seriously affect the perception level, such as bad smell and noise.

From the perspective of visual suggestions, visual perception is not the absolute green visual ratio. Rather, it should be rationally proportioned and optimized according to the site and functional requirements to promote greening and other natural landscapes, which can improve the perception level of the visual senses. In areas dominated by architectural landscapes, it is necessary to increase the green visual rate, while UGBS, located in the urban fringe, needs to increase the number of artificial visual elements and improve service levels and facilities. Noise control is applicable at any site. UGBS generally provides a better sound environment for residents, especially natural sound, which will be beneficial to promote the overall level of perception. Introducing a natural environment with a place to eat or drink for relaxation can boost perception levels, whereas controlling the production of unpleasant smells in the environment can avoid extreme negative emotions. UGBS provides a local low-temperature environment for the city, but overall, it is neglected, suggesting a need to add shade facilities or more vegetation in some places. With the help of sensory maps, we can quickly find good sensory sites and bad sites and maintain or improve them accordingly.

The contribution of this study was to measure the relationship between the physical environment and senses, how they affect the overall perception level, and provide a new method to propose targeted improvement strategies according to different situations and senses. This study takes the lead in using BERT technology to classify and analyze the subjective perception of different senses, which fills the gap of previous studies and provides a new method for the analysis of online reviews. The latest NLP and FCN were used to process text and picture information in online reviews, respectively. Combined with the research objectives, the feature words were divided into the physical environment and sensory descriptions, and a word list was created to quantify the pictures and words.

However, this study had some limitations. Although online reviews can overcome the limitations of field research or laboratory simulations and acquire data not restricted by volunteers’ region, identity, age, or other social identities, the description of some senses is incomplete because of this freedom. Although some senses affect perception, the reviewer may not necessarily express them in the text. At the same time, because Google Maps is used by people with multilingual backgrounds, the current technology may still cause some misunderstanding because BERT is pre-trained in English. In addition, the current picture training database uses the ADE_ 20 K dataset, which has complete picture classification; therefore, it is sufficient to meet the needs of this study. However, in the physical environment, some elements may have identification errors. Overall, at the current technical level, this study has a wide range of reference values and can also be applied to the analysis of public perception in other venues.

Online reviews will not only be in the form of simple text or picture data, which we have not collected at present. There may be an increasing number of reviews, including video and audio, which may also contain a lot of multisensory information and can be analyzed in the future. The development of technology will provide us with the possibility to study a wider range of senses.

Tokyo’s 23 districts have different economic development statuses and degrees of development. The planning and management status of the UGBS can be evaluated based on the number of online reviews, overall perception level, and sensory descriptions. This study used DL to make a quick overall assessment of different sites, which was conducive to a targeted understanding of the public’s subjective perception of them. This study provides a way to quickly acquire and process the public’s subjective perception based on multiple senses through DL, which is helpful to the rapid analysis of specific environments and problems, making UGBS development in different districts more balanced and equitable, promoting sustainable management and development of the site.