Towards Well-Being in Old Residential Areas: How Health-Promoting Environments Influence Resident Sentiment Within the 15-Minute Living Circle

Abstract

Building healthy communities is crucial for creating healthy cities and improving residents’ well-being. Old residential areas, with their substantial stock and elevated health risks, require urgent environmental upgrades. However, the relationship between community health promotion factors and resident sentiment, a crucial indicator of subjective well-being, in old residential areas remains poorly understood. By integrating big data-based community health promotion factors and Weibo data within the 15-min living circle of old residential areas in Xi’an, we developed an XGBoost model and employed SHAP analyses to interpret predictive outcomes. Results show that healthy facilities were dominant influencing factors in old residential areas. Densities of parking, supermarkets, education, package stations, and scenic spots exhibit nonlinear relationships with positive sentiment, indicating clear threshold effects and saturation effects. Two dominant patterns were observed in interactions between dominant factors and their strongest interacting factors. Four environment–sentiment patterns were identified for targeted planning interventions. It is recommended that planners and policymakers account for density phases and synergistic combinations of the dominant factors to optimize community health within old residential areas. The findings offer empirical support and planning insights for fostering healthy, sentiment-sensitive retrofit in old residential areas within the 15-min living circle.

1. Introduction

Rapid global urbanization delivers unprecedented convenience but also intensifies challenges to human well-being and environmental sustainability [1]. In response, the pursuit of “healthy cities” has emerged as a critical global agenda, aiming to shape urban environments that actively promote the health of residents [2]. However, current research and practice have predominantly focused on new or high-end ecological communities as model sites for health-oriented development. In contrast, Old Residential Areas (ORAs), established neighborhoods experiencing a critical process of physical and functional aging in their built environment, form a substantial part of the urban fabric worldwide and house millions of residents, yet have been largely marginalized in the global discourse on community health performance [3]. This oversight represents a significant blind spot, particularly against the backdrop of a growing global emphasis on the regeneration of existing urban stock [4], which positions ORAs as pivotal frontlines for achieving sustainable and healthy urban futures.

Transforming ORAs into health-promoting environments presents a global challenge: the misalignment between conventional retrofit practices and the multifaceted determinants of resident well-being. This challenge is acutely evident in China’s extensive campaign [3,5,6], which has primarily emphasized rectifying physical decay and ensuring basic safety through measures such as roof maintenance and elevator installations [7,8]. Although these engineering-led interventions are necessary, they have largely overlooked the systematic enhancement of the community’s built environment, including the balanced allocation of health-oriented facilities, access to blue-green spaces, and overall streetscape quality. This narrow focus on fundamental repairs has limited the efficacy and broader impact of retrofits, failing to unlock the full potential to deliver comprehensive, higher-order health and well-being benefits. This gap constitutes a dual challenge for both theory and practice: a deficient understanding of the complex relationships linking the built environment to health outcomes. Addressing this imperative, our research aims to model the specific relationships between these key environmental attributes and resident health, thereby generating an evidence base to steer future interventions from basic upgrades toward truly health-centric transformations.

Sentiment serves as a crucial indicator of subjective well-being and psychological satisfaction with community environments [9]. Positive sentiment is indicative of happiness and has been linked to enhanced psychological and physical health [10,11,12]. However, previous sentiment research has predominantly focused on larger scales, such as national [1], urban [10], and district levels, as well as specific areas like urban villages [13], and high-rise residential zones [14]. In contrast, studies examining the impact of built environments on residents’ sentiments in ORAs remain inadequate. Traditional methods employed in sentiment research, including questionnaires [15,16], on-site interviews [15], attention testing [17], and observation techniques [18], are often constrained by small sample sizes, limited geographical coverage, and short temporal spans [19]. These limitations impede the depth and breadth of insights that can be derived, making it challenging to comprehensively elucidate the complex relationship between the built environment and residents’ sentiments. Social media platforms, such as Twitter [20,21], TikTok [22], Facebook [23], Weibo [1], Xiaohongshu [24], have emerged as valuable sources of data. These platforms provide a space for individuals to express opinions and convey emotions. Social media data can capture vast amounts of information in real-time and dynamically, offering a more comprehensive and nuanced understanding of residents’ sentiments [25]. Natural language processing (NLP) technologies, capable of efficiently analyzing large-scale text data and interpreting emotional expressions, provide a robust solution for analyzing social media texts [26,27].

Currently, the mainstream healthy community assessment standards have proposed multi-dimensional frameworks as strategies to enhance community health potential [28]. However, the assessment methods typically rely on community design plans [29,30], planning decisions, documentation [30,31], performance data [31], operational effectiveness [32], and on-site research. These approaches are time-consuming and costly, making it challenging to conduct comprehensive health performance evaluations for many communities within a city. For ORAs, the complexity of the built environment and the lack of systematic planning and design data render these assessment methods even less applicable. Consequently, there remains a significant gap in urban-level analyses of how multiple built environment factors collectively impact health in ORAs. To address this gap, this study capitalizes on urban multi-source big data and Artificial Intelligence (AI) algorithms to efficiently gather and analyze data related to the built environment.

This study takes Xi’an as an example to investigate the relationship between residents’ positive sentiment scores (PSS) and community health promotion factors within the 15-min living circle of ORAs. The 15-min living circle has emerged as a fundamental spatial unit for fulfilling the residents’ basic needs [33], and it is within this immediate geographic scope that environmental factors exert the most direct and significant influence on health. The research methodology consists of three main stages. First, data on health-promoting environment factors will be sourced from multiple big data platforms and processed using ArcGIS and DeepLab V3+. Second, Weibo data will be collected and analyzed using Bidirectional Encoder Representations from Transformers (BERT) to quantify PSS. Third, regression simulation will be conducted using the eXtreme Gradient Boosting (XGBoost) algorithm, with model outputs interpreted by SHapley Additive exPlanations (SHAP) analysis, focusing on feature importance, nonlinear relationships, and interaction effects. Based on SHAP values, distinct typologies of built environment impact were identified, paving the way for data-driven and precision planning strategies. Our research aims to answer the following questions:

(1)

What are the dominant built environment factors that influence residents’ sentiments?

(2)

How do these dominant factors influence residents’ sentiments?

(3)

How do the dominant factors interact with other factors to affect sentiments?

(4)

What types of environment-sentiment patterns exist that can guide targeted planning interventions?

Our findings elucidate built environment factors’ impacts on residents’ sentiments, providing a more comprehensive and scientific foundation for the development of the 15-min living circle of ORAs, offering valuable insights for targeted renovation and health promotion.

2. Literature Review

2.1. Health-Promoting Environment Factors

Early research on the impact of the built environment on human sentiments largely relied on coarse-grained data, such as census statistics [34] and land use classifications [30,31]. Advances in geospatial technologies and big data analytics have now enabled a more refined and multidimensional characterization of built environments. By integrating multi-source data, including OpenStreetMap (OSM), building footprint data, Point of Interest (POIs), remote sensing data, and Street View Images (SVIs), this study synthesizes evidence on key built-environment factors [29,30,31,32,35,36,37,38], categorizing them into four domains: land use intensity, health facility, vegetation and water, and streetscape perception.

2.1.1. Land Use Intensity

Land use intensity domain describes the density and concentration of physical structures within a given area [39]. These factors influence sentiment well-being by shaping residents’ activity patterns, social interactions, and access to amenities [40].

Land use mix of function refers to the integration of different land use types within a neighborhood. It is found that a higher land use mix of functions enhances walkability, reduces car dependency, and improves access to diverse services [41]. These outcomes can foster street life and social interaction, which are potential drivers of positive sentiment. However, excessive functional mixing may also generate negative externalities, such as noise pollution, a lack of tranquility, and environmental disorder, potentially leading to sensory overload and stress [42,43]. Consequently, the sentiment impact is contingent on a balance between vibrancy and calm. Building Coverage Ratio (BCR) is a key metric of urban compactness. Xia et al. found that a higher BCR, indicating a large footprint of buildings, is often correlated with limited personal space, reduced sunlight exposure, and obstructed views, which may evoke feelings of confinement and stress [44]. Street density (SD) reflects the connectivity of the urban road network. Excessively low SD can lead to isolation and poor accessibility, whereas very high density could be frequently associated with increased traffic noise, pedestrian-vehicle conflict, and perceived chaos [45]. An optimal, moderate level of SD supports walkability by facilitating physical activity and social interaction, thereby contributing to sentiment well-being [46]. Old residential unit density often serves as a proxy for neighborhood socioeconomic status and physical deterioration. A high concentration of old housing areas may indicate a lack of public or private investment and substandard living conditions. Exposure to these environmental cues may engender a sense of insecurity and powerlessness among residents, which are established risk factors for negative sentiments.

2.1.2. Healthy Facility

Health Facilities provide environmental support for residents’ sentiment well-being. The accessibility of facilities could meet diverse social needs, reduce daily stressors, and convey a sense of environmental support and safety, thereby promoting sentiments.

Medical facilities deliver critical health security by enabling timely and convenient access to professional medical care, which is vital for patients and older adults. The accessibility of medical facilities could alleviate disease-related anxiety and uncertainty, and serve as a foundation for long-term mental health [47]. Sports facilities and parks and plazas serve as key venues for physical activity and nature contact [48]. Sports facilities support exercise as one of the most effective behavioral interventions for mitigating stress, anxiety, and depression [49,50]. As public open spaces, parks and plazas offer low-cost opportunities for social interaction, recreation, and nature engagement [51], which help both mental health and physical health [52]. Cultural facilities provide rich cognitive stimulation and aesthetic experiences and foster social inclusion and cultural participation, thereby enhancing positive sentiments [53]. Scenic spot density reflects the concentration of unique natural or cultural resources with inherently strong restorative properties [54]. These attractions encourage exploration and visitation, serving as important sources of pleasurable experience and contributing to both sentiment promotion and stress buffering.

Public transport station and supermarket densities reflect neighborhood convenience [55]. Public transport may reduce commuting time and cost while enhancing mobility autonomy and flexibility [56]. Horton et al. found that an appropriate density of supermarkets could facilitate easy access to fresh food and daily necessities, which is important to decrease high rates of chronic disease, thereby influencing patients and their families’ mental health [57]. Together, they alleviate daily burdens and improve residents’ sense of efficiency. Density of public toilets, package stations, and pet services significantly influences sentiment experiences in specific contexts and for particular groups. Public toilet density is critical for enabling safe outdoor activities for the elderly and children [58]. Its absence may cause anxiety and avoidance behaviors. Package station density reflects logistical convenience in the e-commerce era, reducing the hassle of package retrieval and aligning with contemporary lifestyles [59]. Pet service density supports pet owners’ care needs, and since pet ownership is known to reduce stress and enhance social ties, accessible services can amplify these positive effects. Parking density addresses the shortage of parking space in ORAs, mitigating related stress for residents [60]. However, high parking density often indicates car-dominated areas, which may be accompanied by traffic congestion, noise, air pollution, and reduced pedestrian space [61]. These conditions can undermine pedestrians’ sense of safety and comfort, thereby increasing irritability and stress.

Overall, the diversity and accessibility of health facilities help construct a supportive and convenient daily environment that meets residents’ complex needs [62]. By reducing stressors and enhancing positive resources, they collectively protect and promote sentiment well-being.

2.1.3. Vegetation and Water

Natural environmental elements influence sentiments through humans’ innate affinity for nature, with remote sensing enabling effective quantification of these macro-scale ecological factors.

Normalized Difference Vegetation Index (NDVI), a key measure of green vegetation coverage and vitality, is widely used as a proxy for nature exposure in sentiment studies [63]. Natural settings allow directed attention mechanisms to recover and reduce negative sentiments [64]. For example, Wang et al. found that a higher NDVI could lower stress, reduce depression risk, and greater subjective well-being [65]. Normalized Difference Water Index (NDWI) identifies surface water distribution. Blue spaces offer restorative benefits comparable to green spaces [66]. Aquatic environments provide distinctive aesthetic experiences and sensory pleasure. For example, Poulsen et al. indicate that visual exposure to water can induce calming effects and offer restorative benefits [67]. Land Surface Temperature (LST) directly captures urban thermal conditions and serves as a core indicator of the urban heat island effect [68]. High LST not only causes physical thermal discomfort, but also indirectly impairs mental health by limiting outdoor activity and social engagement [1].

2.1.4. Streetscape Perception

Streetscape perception derived from street view images quantifies immediate environmental exposure at the human scale, offering a refined perspective on environment-sentiment relationships [69].

Greenness, measured from a pedestrian viewpoint, quantifies visible vegetation by capturing nature exposure during daily activities [70]. Openness, often indicated by sky view factor, is associated with unobstructed views and correlates with reduced oppression and greater perceived safety, thereby promoting sentiment relaxation [71]. Conversely, crowdedness reflects excessive visual density of buildings, vehicles, or pedestrians, acting as an environmental stressor that may trigger anxiety and avoidance behavior [72]. Imageability and walkability describe aesthetic appeal and functional quality. Imageability refers to a place’s distinctiveness and ability to evoke curiosity and attachment, enhancing cognitive engagement and positive sentiment [73]. Walkability integrates safety, comfort, and interest. Highly walkable environments encourage physical activity [74] and informal social interaction, both key contributors to sentiment well-being. Streetlight view index specifically addresses perceived safety at night. Adequate street lighting serves as a critical cue for security after dark, reducing fear of crime and promoting nighttime mobility and socialization [75]. By extending opportunities for positive activity into night hours, it contributes to holistic sentiment health.

2.2. Methods for Measuring Sentiments

Sentiment analysis aims to computationally identify, extract, and quantify subjective emotional information from text [76]. In the early stages of research, traditional word embedding methods such as Word2Vec and GloVe served as mainstream techniques for measuring sentiment in text. However, these methods exhibit notable limitations. They produce static word vectors, meaning each word is represented by a fixed vector regardless of its contextual usage [77]. This assumption severely restricts their ability to handle polysemy. Furthermore, the architectural simplicity of these models, often implemented using shallow neural networks, limits their capacity to capture long-range bidirectional contextual information [78]. As a result, they often fall short in interpreting complex semantic structures and nuanced emotional expressions.

The advent of deep learning marked a new era in sentiment measurement technologies. Models such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), and Convolutional Neural Networks (CNNs) enabled the automatic learning of hierarchical textual features, overcoming the limitations of manual feature engineering inherent in earlier approaches [79]. Nevertheless, the true paradigm shift came with the introduction of pre-trained language models based on the Transformer architecture, including BERT, A Robustly Optimized BERT Pretraining Approach (RoBERTa), and the Generative Pre-Trained Transformer (GPT) series. Among these, BERT has played a particularly important role. Its core strengths lie in its deep bidirectional encoder and multi-head self-attention mechanism, which allow it to process entire sequences in parallel and dynamically capture complex contextual dependencies between words [80]. This architecture enables BERT to generate dynamic contextualized word representations, effectively modeling polysemy [81].

2.3. Methods for Measuring Streetscape Perception

The early stages of streetscape perception research were marked by limitations in both data collection and analytical methods. In terms of data collection, studies primarily relied on field-based approaches, such as in-person interviews, questionnaires, and basic photographic documentation [70]. These methods were difficult to scale and offered limited reproducibility [72]. On the analytical front, interpretation depended heavily on researchers’ or limited experimenters’ subjective observations of streetscape features, leading to considerable personal bias, low efficiency, and poor reproducibility.

Advances in geographic information systems and internet technologies have laid the foundation for large-scale streetscape data collection and processing [72]. Emerging acquisition technologies, such as vehicle-based mobile mapping systems and unmanned aerial vehicles, have been used to auto capture the high-resolution, multi-perspective streetscape imagery [82]. These developments facilitated the establishment of panoramic mapping services such as Google Street View and Baidu Street View, which have also generated easily accessible databases of SVIs [69]. This has provided unprecedented data support for urban environmental studies.

In terms of analytical methods, progress in deep learning-based computer vision has substantially enhanced the accuracy and scalability of streetscape image analysis. Early semantic segmentation methods largely relied on low-level visual features such as color and texture, which offered limited discriminative power, resulting in poor robustness and accuracy [83]. In contrast, modern semantic segmentation models leverage deep neural networks to automatically learn complex mappings from pixels to semantic categories. Models like Fully Convolutional Networks (FCN), U-Net, and the DeepLab series extract highly discriminative multi-level features, enabling precise pixel-wise classification and significantly improving the accuracy and robustness of streetscape element recognition. Among these, DeepLab V3+ stands out for its exceptional performance in complex scenarios. By integrating an Atrous Spatial Pyramid Pooling module, it captures multi-scale contextual information effectively, while its encoder–decoder architecture enables precise boundary recovery [72]. These attributes make it particularly suitable for applications demanding high segmentation detail [84], such as fine-grained urban planning.

2.4. Methods for Data Analysis

In early research on the relationship between the built environment and sentiments, traditional linear regression models served as the dominant analytical paradigm [85]. However, these models suffer from inherent limitations. Their core assumption is that a simple linear relationship exists between independent variables and the dependent variable and that the effects of these variables are mutually independent. But several studies have proved that environment-sentiment relationships are far more complex, often exhibiting significant nonlinear characteristics [86]. Traditional models struggle to capture these complex patterns.

Machine learning algorithms have demonstrated considerable potential in this field of research, leveraging their data-driven nature and inherent capacity for nonlinear modeling [87]. Unlike parametric models that rely on strong a priori assumptions, machine learning algorithms can automatically learn and identify complex mapping relationships from data without requiring pre-specified functional forms [88]. This flexibility enables them to effectively uncover hidden nonlinear patterns and interaction effects [89] between built environment variables and sentiment outcomes, thereby providing theoretical insights that more closely reflect real-world complexities and yielding more accurate predictive performance. The XGBoost model has gained prominence due to its exceptional performance. Compared to other decision tree-based algorithms, XGBoost incorporates regularization terms into the loss function and employs refined techniques in gradient optimization, which enhance predictive accuracy and effectively mitigate overfitting [87]. Furthermore, the SHAP framework, rooted in cooperative game theory based on Shapley values, provides a powerful tool for interpreting the “black-box” nature of complex models like XGBoost [90,91]. SHAP quantifies the marginal contribution of each built environment feature to the model’s final prediction and presents it as an attribution value [89,92]. This enables researchers not only to develop high-accuracy predictive models but also to clearly interpret how and to what extent each variable influences sentiment outcomes. Such interpretability offers a robust scientific foundation for formulating targeted urban health intervention strategies.

3. Materials and Methods

3.1. Research Framework

The research framework adopted in this paper is illustrated in Figure 1. First, this study defined the 15-min walking living circle of old residential areas as the study area. Second, Appropriate urban big-data sources were selected to characterize health-promoting environment factors, including OSM, POIs, building footprint data, SVIs, and Landsat data. Weibo check-in data, with its precise geographic location tags [10], was used to analyze the dependent variable residents’ sentiments. Third, NLP techniques were applied to perform sentiment analysis on “check-in” data. DeepLab V3+ was utilized for the semantic segmentation of SVIs. All the variables were processed and integrated in ArcGIS 10.8.1 and Excel 16.84. Fourth, this study employed the XGBoost algorithm for regression simulation. Fifth, SHAP analysis was used to explain the XGBoost model, specifically exploring the significance of variables, non-linear relationships, and interaction effects. Based on SHAP values, distinct typologies of built environment impact were identified using K-Means Clustering (KMeans).

3.2. Study Area

Xi’an, the capital of Shaanxi Province in northwestern China, has the longest history among the six ancient capitals of China [93]. This study focuses on the six central urban districts: Beilin, Yanta, Lianhu, Baqiao, Xincheng, and Weiyang. As the core urban area of a historically rich and densely populated metropolis, this region contains a high concentration of ORAs, making it an ideal case for this research.

This study employed Python (version 3.8) web crawler technology to obtain data on ORAs constructed in and before 2004 within the six central urban districts of Xi’an from the Anjuke website, one of China’s leading real estate transaction platforms. The initial data collection yielded 1779 ORAs. Based on the POIs of these ORAs, 15-min living circles were constructed based on the road network in ArcGIS. To ensure sufficient data volume and avoid numerical bias for subsequent sentiment analysis, ORAs with fewer than 100 sentiment data points in their 15-min living circle were excluded. The final 15-min living circle for the study was established again based on the road network in ArcGIS. The study area was divided into 100 m × 100 m grids as statistical units for all subsequent spatial analysis.

3.3. Variables

Based on the literature review in Section 2, this study established a system of community health-promoting factors. This system comprises four categories: land use intensity, health facility, vegetation and water, and streetscape perception, totaling 25 indicators. This section provides details on the data source and calculation formulas for these factors.

Table 1. Variables system.

Category	Domains	Variables	Source
Dependent Variable	Sentiment	Positive sentiment score (PPS)	Weibo data
Independent Variable	Land use intensity	Land use mix of function	POIs
		Building coverage ratio (BCR)	OSM and building footprint data
		Street density (SD)
		Old residential unit density	POIs
	Healthy facility	Public transport station density	POIs
		Parking density
		Sports density
		Education density
		Medical density
		Culture facility density
		Park and plaza density
		Scenic spot density
		Supermarket density
		Public toilet density
		Package station density
		Pet service density
	Vegetation and water	Normalized Difference Water Index (NDVI)	Landsat 8
		Normalized Difference Water Index (NDWI)
		Land Surface Temperature (LST)
	Streetscape perception	Greenness	SVIs
		Crowdedness
		Imageability
		Openness
		Walkability
		Streetlight view index

displays the variables and their data sources. All factors were using a grid as the spatial unit of analysis. Factors in the land use intensity domain and the health facility domain were calculated in ArcGIS. Factors in the vegetation and water domain were calculated using ENVI 5.3. Deep Learning algorithm DeepLab V3+ was utilized for the semantic segmentation of SVIs. The BERT model was used to perform sentiment analysis. All the calculation results were integrated into ArcGIS and Excel.

3.3.1. Sentiment

PSS refers to the average positive sentiment score in each grid [19]. The formula is shown below:

$${PSS}_i={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{SS}_i$$

(1)

where ${SS}_i$ denotes the positive sentiment score of one text, and $n$ denotes the quantity in one grid.

3.3.2. Land Use Intensity

• Land use mix of function

The Land use mix of function (measured by the Shonnon-Weaver Diversity) refers to the degree of integration of different types of land uses (such as residential, commercial, and industrial) within each grid. The formula [94] is shown below:

$$SHDI=-\sum _{i=1}^m{p}_i\times \ln p_i$$

(2)

where $SHDI$ denotes the Shonnon-Weaver Diversity, $p_i$ denotes the proportion of land use type $i$ in all types of POIs, $m$ denotes the total number of land use types.

2.

Building coverage ratio

BCR is the ratio of the total base area of all buildings to the total area of the grid [19]. The formula is shown below:

$$BCR={\displaystyle \frac{B}{A}}$$

(3)

where $B$ and $A$ denote the building base area and grid area.

3.

Street density

SD is the length of roads per grid area (km/km²). The formula is shown below:

$$SD={\displaystyle \frac{L}{RA}}$$

(4)

where $L$ denotes the length of roads, $RA$ denotes the grid area.

4.

Old residential unit density

Old residential unit density refers to the kernel density of old residential units of grid area (counts/km²). Its calculation uses the same method as health facility density, which is presented below.

3.3.3. Healthy Facility

The series of facility density, including public transport station density, parking density, sports density, education density, medical density, culture facility density, park and plaza density, scenic spot density, supermarket density, public toilet density, package station density, pet service density, as well as old residential unit density are calculated using this unified method, which refer to the kernel density of each healthy facility of grid area (counts/km²). The formula [95] of Kernel density estimation is as follows:

$$f(x)={\displaystyle \frac{1}{nh}}\sum _{i=1}^{n}K({\displaystyle \frac{x-x_i}{h}})$$

(5)

where $f\left(x\right)$ represents the density estimate at location $x$, $n$ denotes the total number of events, $K$ serves as a kernel function that contributes to the estimated position $x$ density of each event, $x_i$ corresponds to the location of each event, and $h$ is the radius.

3.3.4. Vegetation and Water

• NDVI

NDVI reflects the vegetation coverage density and growth status within each grid. The formula is shown below:

$$NDVI=(NIR-Red)/\left(NIR+Red\right)$$

(6)

where $NIR$ represents the reflectance value of the near-infrared band, selecting Band 5, $Red$ represents the red band reflectance value, selecting Band 4.

2.

NDWI

NDWI is defined as the intensity of water distribution within a grid, serving as a proxy indicator for residents’ exposure to “blue spaces”. The formula is shown below:

$$NDWI=\left(GREEN-NIR\right)/\left(GREEN+NIR\right)$$

(7)

where $GREEN$ represents the green band, selecting Band 3.

3.

LST

The LST data from Landsat 8 TIRS Collection 2 Level 2 science product has already been processed with atmospheric correction and provides surface temperature values in Kelvin. The pixel values were converted from Kelvin to Celsius, and the mean value within each grid was computed to serve as its thermal environment indicator. The formula is shown below:

$$LST\left(°C\right)=\mathrm{S}\mathrm{T}\_\mathrm{B}10-273.15$$

(8)

where $\mathrm{S}\mathrm{T}\_\mathrm{B}10$ denotes the pixel value from Landsat 8 Band 10, representing the surface brightness temperature in Kelvin (K).

3.3.5. Streetscape Perception

• Greenness

Greenness refers to the vegetation coverage rate observed from a human perspective. The formula is shown below:

$$G_i={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{T}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{P}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{G}_i$$

(9)

where $T_i$ denotes the proportion of tree pixels, $P_i$ denotes the proportion of plant pixels, $G_i$ denotes the proportion of grass pixels. $n$ denotes the number of images.

2.

Crowdedness

Crowdedness reflects the degree of congestion and the intensity of activity constituted by pedestrians and vehicles in the street environment. The formula is shown below:

$$C_i={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{Ps}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{C}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{Bs}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{Tr}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{V}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{MB}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{Bk}_i$$

(10)

where ${Ps}_i$ denotes the proportion of person pixels, $C_i$ denotes the proportion of car pixels, ${Bs}_i$ denotes the proportion of bus pixels, ${Tr}_i$ denotes the proportion of truck pixels, $V_i$ denotes the proportion of van pixels, ${MB}_i$ denotes the proportion of motorbike pixels, ${Bk}_i$ denotes the proportion of bike pixels.

3.

Imageability

Imageability is the unique quality of a place, endowing it with distinctiveness, recognizability, and memorability [70]. The formula is shown below:

$$I_i={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{B}_i+{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{S1}_i$$

(11)

where $B_i$ denotes the proportion of building pixels, ${S1}_i$ denotes the proportion of signboard pixels.

4.

Openness

Openness reflects the visibility of the sky in streets [10]. The formula is shown below:

$$O_i={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{Sky}_i$$

(12)

where ${Sky}_i$ denotes the proportion of sky pixels.

5.

Walkability

Walkability reflects the level of street support for outdoor walking [70]. The formula is shown below:

$$W_i={\displaystyle \frac{\frac{1}{n}{\sum }_{i=1}^n{P1}_i+\frac{1}{n}{\sum }_{i=1}^n{F}_i}{\frac{1}{n}{\sum }_{i=1}^n{R}_i}}$$

(13)

where ${P1}_i$ denotes the proportion of pavement pixels, $F_i$ denotes the proportion of fence pixels, $R_i$ denotes the proportion of road pixels.

6.

Streetlight view index

Streetlight view index indicates the visibility of streetlights in the streetscape, reflecting the brightness of the road at night and, to some extent, the safety of the road at night. The formula is shown below:

$${SL}_i={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{Sl}_i$$

(14)

where ${Sl}_i$ denotes the streetlight pixels.

3.4. Data Collection

The ORA data was collected from Anjuke.com (https://xa.anjuke.com/community/ (accessed on 26 July 2024)). The data includes name, location, completion time, the total number of households, building type, floor area ratio, ownership type, green ratio, average selling price, etc.

Weibo “check-in” data in Xi’an City from May 2024 to October 2024 was retrieved by Weibo API, including latitude and longitude, location, post text, post time, etc. The geographic locations of these data were obtained officially through the POI coordinates manually selected by the user, or through geolocation permission, which uses GPS, cellular network station, or Wi-Fi coordinates [96]. A total of 179,380 texts were gathered.

The building footprint data, including building outline vectors (https://doi.org/10.5281/zenodo.8174931 (accessed on 18 November 2024)), was obtained from Zenodo [97]. The Google Earth imagery data used for building identification in this dataset was collected between 2020 and 2022. The data was produced within a large-scale mapping framework and is of high quality, with a spatial resolution of 0.5 m.

The street data was obtained from the Open Street Map (https://www.openstreetmap.org (accessed on 2 October 2024)). OSM provides global-scale geographic data, including spatial coordinates of buildings at different levels, basic outlines, floors, streets, and more.

POI data was collected from Baidu Maps (https://map.baidu.com/ (accessed on 25 September 2024)), including name, address, latitude and longitude of location, categories, etc. There were initially 23 categories of data, including catering services, scenic spots, companies, shopping, transportation facilities, financial and insurance services, etc. After filtering, integrating, and reclassifying the data, a total of 229,540 POIs were obtained for the calculation of the land use mix of function, and 37,531 POIs of healthy facilities were obtained.

NDVI, NDWI, and LST were all based on the Landsat 8 Collection 2 Level 2 [98], which were obtained from https://earthexplorer.usgs.gov (accessed on 10 November 2024).

SVIs were obtained by the Baidu Street View API (https://lbsyun.baidu.com (accessed on 16 November 2024)). The SVI sampling dot was conducted along the street with 50-m intervals in the ArcGIS. The 360$°$ panoramic photo was collected at each sampling point. A total of 24,918 photos were gathered.

3.5. Data Processing

3.5.1. Sentiment Analysis Using BERT

Data cleaning is employed to identify and rectify errors within the dataset. Data entries lacking or containing missing latitude and longitude information are removed to eliminate invalid data. The data is then deduplicated based on Weibo links. Geospatial filtering is applied to select Weibo data within the study area. Subsequently, the data is further refined using keywords: old residential unit, old residential area, old neighborhood, old housing, community, aged community, micro-renovation, renovation, and 15-min living circle. The selected Weibo texts are then subjected to sentiment prediction using the BERT, yielding sentiment polarity and sentiment scores. The score reflects the probability tendency or state of sentiment [13], indicating the degree of residents’ sentiments towards the living environment.

BERT is a pre-trained language model released by Google in 2018. This algorithm demonstrates higher accuracy and better performance in context awareness and semantic capture [81]. In this study, the BERT-based Chinese model is utilized. This model is a fine-tuned BERT model designed for the Chinese language. The “Weibo_senti_100k” dataset, which contains over 110,000 Weibo comments labeled with positive or negative sentiment, is used as the pre-training dataset [99].

First, the Weibo text preprocessing is encapsulated using the WeiboPreprocesser class, which includes text cleaning, specific text deletion, and text filtering. Text cleaning involves removing HTML tags, replacing URLs with [URL], replacing @usernames with [user], removing topic tags, eliminating content within square brackets, normalizing whitespace, replacing emoticons, and retaining Chinese punctuation while converting full-width characters to half-width. Specific text deletion identifies and removes Weibo posts containing certain patterns that are web structural or platform tracking information, which are irrelevant to sentiment analysis. Text filtering eliminates overly short texts or those that do not contain Chinese characters. Subsequently, the BertTokenizer is employed via the tokenize_for_bert method to encode the text into the input format required by the BERT model. The data is then randomly split into training (80%) and testing (20%) sets. Model parameters are set, and early stopping is implemented to prevent overfitting. Model performance is evaluated using accuracy, precision, recall, F1 score, and confusion matrix. Once the performance meets the criteria, final predictions are made.

3.5.2. Semantic Segmentation Using DeepLab V3+

In this study, we employed DeepLab V3+ for the semantic segmentation of street view images to identify street elements. Subsequently, street features were quantified to characterize street perception.

We utilized a pre-trained DeepLab V3+ model, namely deeplab_resnest101_ade, in this study. This model, available in the GluonCV library, is based on the DeepLab V3+ architecture with ResNeSt101 as the backbone network and has been pre-trained on the ADE20K dataset. The ADE20K dataset is an open-source semantic segmentation dataset released by the CSAIL Vision team at MIT, comprising 27,574 annotated scene images across 150 categories [100,101].

First, images were read using mxnet.image.imread. Second, the test_transform was applied to convert the images into the format required for model input, involving normalization, channel order adjustment, and resizing. The deeplab_resnest101_ade model was then employed for image prediction. The category labels for each pixel were obtained using mx.nd.argmax, yielding the segmentation results. The model output includes the visualization of the segmentation results (Figure 2) and a table of segmentation category proportions.

This study gathered tree, plant, grass, sky, building, fence, road, sidewalk, streetlight, person, truck, car, van, bike, motorbike, bus, and signboard element pixel proportions and used them to calculate the streetscape perception factors.

3.6. Data Analysis

The XGBoost algorithm was employed for data analysis. The dataset was randomly split into the training and testing sets at an 8:2 ratio. First, a grid search with 5-fold nested cross-validation (CV) was employed to optimize the hyperparameter combination. Second, the training set was evaluated using 10-fold cross-validation (xgb.cv) with a maximum of 20,000 boosting rounds, incorporating early stopping to mitigate overfitting and enhance generalization. The mean and standard deviation of training and validation root mean squared error (RMSE) were tracked across iterations to assess model stability, with the optimal iteration count (best_rounds) determined by early stopping. Third, the final model was trained on the complete training set using the tuned hyperparameters, where n_estimators was set to best_rounds. Model performance was quantified on both training and test sets using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and the coefficient of determination (R²).

3.7. Outcomes

3.7.1. Model Interpretability: SHAP Analysis

SHAP analysis was applied to uncover the underlying drivers of the model’s output by decomposing the prediction into contributions from each explanatory variable.

First, TreeExplainer was employed to compute SHAP values for each sample in the testing set. Global feature importance rankings were determined by calculating the mean absolute SHAP values, with visualizations including feature importance bar plots and feature distribution beeswarm plots. For deeper analysis, the diagonal elements of the SHAP interaction value matrix were extracted as main effect values, representing features’ independent contributions to predictions [88]. For the top 12 most important features, SHAP main effect value scatter plots were generated to reveal nonlinear relationships between feature values and model predictions. Then, locally Weighted Scatterplot Smoothing (LOWESS) was employed to smooth scatter plots and capture the non-linear trends, which is a non-parametric regression technique using the local weighted least squares method [53]. Furthermore, feature interaction effects were quantified using the off-diagonal elements of the interaction value matrix. By computing the mean absolute interaction values between the top 12 features and all other features, the strongest interacting pairs for dominant features were identified. Bivariate interaction plots were generated to elucidate patterns of synergistic or competitive effects between features [88,89].

3.7.2. Cluster Analysis: KMeans

K-means is a widely used unsupervised learning method that partitions data into K clusters by iteratively minimizing the within-cluster sum of squares, with each cluster represented by its centroid. This approach groups samples with similar driver characteristics into representative typologies. Analyzing the internal features of these clusters by examining typical samples reveals the key factors and their effect directions that characterize different patterns of built environment impact on sentiment scores.

The SHAP value matrix, derived from the shap.Explainer, was first standardized via Z-score normalization. The optimal number of clusters (K) was determined using the elbow method (KneeLocator) within a range of 1 to 10. After performing K-means clustering, the most representative samples for each cluster were identified by selecting the five samples with the smallest Euclidean distance to their respective centroid. The SHAP vectors of these typical samples were then visualized using force plots (shap.force_plot) to illustrate the distinct driver profiles for each typology.

4. Results

4.1. 15-Min Living Circle of the ORAs

Following the data screening process detailed in Section 3.2, a final set of 1401 ORAs was retained for analysis, representing 78.7% of the initially collected dataset. The spatial extent formed by their 15-min living circles defines the ultimate geographical scope of this study (Figure 3). Furthermore, this final study area was divided into 17,408 grids of 100 m × 100 m, which served as the uniform statistical units for all subsequent spatial metric calculations.

4.2. Positive Sentiment Score in the 15-Min Living Circle of the ORAs

The BERT model achieved high accuracy (0.9773), precision (0.9938), recall (0.9606), and F1 score (0.9769). These results indicate that the model accurately identified the relevant sentimental values and captured a substantial proportion of the actual sentimental values.

The positive sentiment scores of social media posts were quantified, ranging from 0 to 1. A score closer to 1 indicates a stronger positive sentiment, while a score closer to 0 signifies a weaker positive sentiment and a stronger negative sentiment. The results show that the overall average score of positive sentiment for the 15-min living circle in the ORAs of Xi’an is 0.5575, indicating residents’ overall perception leans slightly towards positive sentiment.

The spatial distribution (Figure 4) reveals two prominent high-positive-sentiment clusters: the Qujiang Cultural Tourism District (southeastern part of the study area) and the contiguous area encompassing the Han City Lake Water Conservancy Scenic Spot and Fengcheng 5th Road Metro Station (northwestern part of the study area). Four negative-sentiment agglomerations are observed: (1) the Tumen district (western part of the study area), (2) the north side district of the Daming Palace National Heritage Park (northeastern part of the study area), (3) the Wanshou Road—Changlepo Metro Station contiguous area and the area east of the Fangzhicheng (eastern part of the study area), (4) the southern Ganjiazhai area (southwestern part of the study area). These regions exhibit consistently higher negative sentiment scores compared to surrounding areas, forming significant sentimental trough zones in the urban landscape.

4.3. Feature Importance Ranking of Community Health Promotion Factors

Figure 5 illustrates the changes in the mean Root Mean Square Error (RMSE) of the XGBoost regression model during the training and testing processes. The decline in the RMSE mean on the training set indicates an enhanced fitting capability of the model on the training data, while the RMSE mean reduction on the test set reflects an improved generalization ability of the model [102]. When both the training and test set RMSE means decrease synchronously and stabilize, it signifies that the model’s generalization ability and stability are enhanced simultaneously.

The best mean RMSE on the test set is 0.062966. The Normalized RMSE (NRMSE), which is the ratio of RMSE to the range of the target variable (the difference between the maximum and minimum values, which is 1 in this study) [102], is less than 10%. This indicates that the prediction error is relatively small compared to the data fluctuation range. This suggests the model has high precision and prediction accuracy.

The evaluation metrics indicate that the errors of the XGBoost regression model are relatively small in both the training and testing sets (

Table 2. Evaluation metrics of the model.

	MSE	RMSE	MAE	MAPE	R2
Training set	0.001571	0.039638	0.027216	0.069255	0.934281
Testing set	0.003441	0.058658	0.039155	0.119638	0.853763

). The testing set’s R² of 0.853763 and MAPE of 0.119638 indicate a satisfactory model performance.

Feature importance ranking of the community health promotion factors affecting the PSS within the 15-min living circle of the ORAs (Figure 6) reveals prominent variations across domains: healthy facilities demonstrate the most significant impacts, followed by land use intensity and vegetation and water, while streetscape perception shows relatively minimal influence.

The top 13 features contribute substantially more to sentiment prediction than other factors. These key features are predominantly (12/13) healthy facilities. Specifically, parking facility density exhibits significantly higher SHAP values than other facilities, followed by supermarket (representing food facilities), education, package station, and culture facility densities. Old residential unit density is the sole land use intensity factor appearing in the top features.

4.4. The Nonlinear Relationships Between the Dominant Factors and PSS

The parking density (Figure 7a) exhibits an overall positive correlation with the PSS. When the kernel density exceeds 57.67, the impact on the PSS shifts from inhibition to promotion. As the density increases, the promoting effect continues to increase before stabilizing. The densities of the supermarket (Figure 7c) and cultural facility (Figure 7f) also show generally positive correlations with the PSS. However, when the density of these facilities surpasses their respective thresholds, the promoting impact is small. The supermarket density exhibits fluctuations during the promotion process.

The densities of the old residential unit (Figure 7b) and sports facility (Figure 7i) exhibit overall negative correlations with the PSS. When the facility densities exceed their respective thresholds, their impact on the PSS shifts from promotion to inhibition. Further increases in density intensify the inhibition impact until stabilization.

The densities of the education (Figure 7d) and package station (Figure 7e) exhibit a “U-shaped pattern”. When the facility density is below the first threshold and above the second threshold, its impact on the PSS is promotion. When the facility density is between the two thresholds, it exerts an inhibitory effect. The medical density (Figure 7h) shows a similar trend but with a minimal impact, hovering close to zero. The public toilet density (Figure 7k) forms an “inverted U shape pattern”. When the facility density is below the first threshold and above the second threshold, it prohibits the PSS. When it is between the two thresholds, its impact is promotion.

The scenic spot density (Figure 7g) consistently exerts a promoting impact on the PSS, with SHAP main effect values fluctuating around 0.01, indicating a relatively stable effect.

The densities of the public transport station (Figure 7j) and park and plaza (Figure 7l) exhibit four-stage patterns. Facility densities have promotional impacts on the PSS when they are, respectively, below the first threshold and between the second and third thresholds. Facility densities have inhibitory impacts on the PSS when they are between the first and second thresholds and above the third threshold. The threshold intervals for public transport stations density are relatively large and evenly spaced. The first threshold for park and plaza is very small, and the first and second thresholds are very close to each other compared to the overall range of values, which warrants further exploration in practical applications.

4.5. Interaction Effects Between the Dominant Factors and Their Strongest Interacting Factors

The SHAP interaction dependence plot quantifies the interaction effects between the top twelve important factors and their strongest interacting factors. The X-axis represents the value of the dominant factor, while the Y-axis represents the SHAP interaction value. The color indicates the value of the strongest interacting factor.

The interaction effects of parking density (Figure 8a) with old residential unit density, old residential unit density (Figure 8b) with parking density, cultural facility density (Figure 8f) with public toilet density, public transport station density (Figure 8j) with old residential unit density, and park and plaza density (Figure 8l) with public toilet density exhibit the same pattern (hereinafter referred to as Pattern A): before the threshold, the dominant factor show a positive (negative) interaction with low (high) values of the strongest interacting factor, indicating a synergistic (competitive) relationship, which enhances (offsets) the sentiment prediction. Beyond the threshold, the interaction reverses: the dominant factor shows a positive (negative) interaction with high (low) values of the strongest interacting factor. Under this pattern, the combination of low (high) values of the dominant factor and low (high) values of the strongest interacting factor will promote the PSS.

The interaction effects of supermarket density (Figure 8c) with scenic spot density, education density (Figure 8d) with package station density, package station density (Figure 8e) with education density, scenic spot density(Figure 8g) with supermarket density, medical density (Figure 8h) with parking density, and public toilet density (Figure 8k) with scenic spot density share a similar pattern (hereinafter referred to as Pattern B): before the threshold, the dominant factor shows a positive (negative) interaction with high (low) values of the strongest interacting factor, indicating a synergistic (competitive) relationship, which enhances (offsets) the sentiment prediction. Beyond the threshold, the interaction reversed: the dominant factor shows a positive (negative) interaction with low (high) values of the strongest interacting factor. Under this pattern, the combination of low (high) values of the dominant factor and high (low) values of the strongest interacting factor will promote the PSS.

The interaction effect of sports density (Figure 8i) with parking density has a four-stage pattern. The first and second stages are the same as Pattern B. The third and fourth stages repeated Pattern B. When facility density is within the first and third stages (the second and fourth stages), a combination with the high (low) value of the strongest interacting factor will promote the PSS.

4.6. Built Environment—Sentiment Patterns Classification

Four (K = 4) typical types of the built environment impact were identified through cluster analysis. Figure 9a illustrates the distribution of four clusters of the impact pattern of the community health promotion factors on residents’ sentiments in the study area. It can be known that Cluster 1 (Figure 9b) is the Positive Sentiment—Comprehensive Health Facilities Promotion Type. The prediction result of sentiments is positive. Almost all health facility density factors have made positive contributions (indicated in red). Among them, the density of education, package station, public toilet, parking, and medical facilities are the factors with the greatest impact. Cluster 3 (Figure 9d) is the Positive Sentiment—Partial Health Facilities Promotion Type. The prediction result of sentiments is positive. The densities of parking facilities, supermarkets, and cultural facilities are the core positive driving factors, while the density of old residential units and educational facilities are the core negative (indicated in blue) driving factors.

Cluster 2 (Figure 9c) is the Negative Sentiment—Insufficient Medical Facilities Dominant Type. The prediction result of sentiments is negative. The densities of medical facilities, supermarkets, scenic spots, and parking facilities are the core negative driving factors, while the density of sports facilities and package stations are the main positive driving factors. Cluster 4 (Figure 9e) is the Negative Sentiment—Insufficient Parking Facilities Dominant Type. The prediction result of sentiments is negative. Almost all health facility densities have a negative impact, with the most prominent factors being the density of old residential areas, parking facilities, and parks and plazas.

5. Discussion

5.1. Prioritization in ORAs Retrofit for Sentiment Promotion

The findings demonstrate that healthy facilities [47,56,103,104,105,106] exhibit the most pronounced influence on the PSS, while the impacts of land use intensity, vegetation and water, and streetscape perception are relatively minor. This finding is in alignment with previous studies that demonstrated that the optimization of public services can significantly enhance subjective well-being [107,108].

Parking density emerges as the most significant factor affecting PSS. Parking scarcity in ORAs has long been a critical issue [109]. The growing mismatch between surging private vehicle ownership and inadequate historical planning has exacerbated parking-related stressors, making it a key determinant of mental health [60]. In areas with insufficient parking supply, residents frequently face difficulties in finding spaces, increased time costs, and reduced travel efficiency [61,109]. Moreover, haphazard parking not only encroaches on public activity spaces but also fuels neighborhood disputes, further degrading community livability and fostering chronic anxiety and irritability, thereby diminishing life satisfaction. This finding confirms that well-planned parking facilities can significantly enhance sentimental well-being [61]. Old residential unit density is the second important factor. The possible reason is that it not only reflects construction characteristics and neighborhood environment but also is closely related to the population size, all of which jointly influence the PSS. Education density significantly impacts PSS. This is mainly because ORAs are in urban centers with high-quality schools [110], attracting homebuyers seeking school district homes. Thus, education density directly affects the fulfillment of residents’ educational needs. An appropriate educational density can reduce residents’ commuting time and costs, enhance the cultural atmosphere and overall quality of the community, further elevating residents’ sense of well-being.

5.2. Exploration of the Nonlinear Relationship Between Community Health Promotion Factors and PSS

These results contribute to the development of a healthy community by demonstrating that the influence of community health promotion factors on the PSS is not linear but rather contingent on threshold effects and saturation effects.

Parking, supermarket, and cultural facility densities initially suppress and then enhance PSS as density increases with distinct thresholds. However, as the density continues to increase, the marginal gains plateau, indicating that excessive infrastructure offers limited sentimental returns. Take parking density as an example, when the kernel density value exceeds 57.67, sentimental benefits rise markedly, whereas under threshold conditions demonstrate an inhibitory effect on sentiment. However, beyond a kernel density of 88, the marginal gains plateau, indicating that excessive parking infrastructure offers limited sentimental returns. These findings carry important implications for ORAs regeneration: the areas below the 57.67 threshold should implement compensatory strategies through shared parking systems and multi-level garages to alleviate parking stress and associated conflicts. The observed plateau effect beyond 88 kernel density suggests policymakers should implement density ceilings in saturated zones, redirect surplus resources to higher return on investment community improvements, and mandate cost-effectiveness analyses for new parking proposals.

Education and package station density show “U-shaped” relationships. This is likely because at low densities, facilities achieve optimal demand-supply matching that reduces resource waste and enhances resident satisfaction. When facility kernel densities reach intermediate levels without achieving network synergy, facility clustering generates negative externalities, such as space encroachment (school-related parking occupying roadways during peak hours) and activity displacement (package cabinets reducing recreational space), which trigger community order complaints and depress PSS. When the facility density beyond the thresholds, facilities generate network effects and labeling effects [111,112]. Education facility clustering forms higher education districts or integrated “preschool-tutoring-extracurricular” ecosystems, while high-density package stations evolve into logistics hubs where services enhance regional delivery quality. When communities acquire positive labels like “education-convenient zones” or “efficient-delivery areas”, residents’ positive sentiments are boosted through enhanced belongingness and convenience. Based on the above findings, priority should be given to eliminating facility “coverage blind spots” to meet basic needs. When facility density reaches the medium density stage, spatiotemporal sharing strategies and dynamic space partitioning should be implemented to alleviate spatial conflicts, while facility layouts should be optimized through community consultation mechanisms, such as staggered school pickup times and centralized delivery nodes. At the high-density stage, monitoring inflection points in sentimental returns from facilities can prevent “oversupply fatigue” caused by excessive construction, and guide facility upgrades toward specialization and networking in high-density areas.

The density of scenic spots has a fluctuating yet overall positive impact on residents’ sentiments. In Xi’an, an ancient historical city, scenic spots are primarily fixed-location cultural relics, which minimally impact residents’ daily lives. Additionally, the long-term development of the tourism industry has positively impacted the urban economy and stabilized residents’ acceptance of tourists [113]. Therefore, the density of scenic spots has an overall positive effect on PSS.

5.3. Exploration of the Interaction Effects Between Community Health Promotion Factors

The interaction effects reveal two dominant patterns, Pattern A and Pattern B. Pattern A emerged in interactions of parking density with old residential unit density, cultural facility density with public toilet density, public transport station density with old residential unit density, and park and plaza density with public toilet density. Low-low and high-high combinations of facility density values show synergistic effects, enhancing the PSS. The likely reason is that these facilities are interdependent in terms of functions and the demands they meet [55,114]. For example, as the density of old residential units increases, so does the population, leading to greater demand for parking and public transportation. Therefore, when the density changes in tandem, they can more effectively satisfy residents’ demands.

Pattern B was observed in the interactions between supermarket density and scenic spot density, education density and package station density, and medical density and parking density. This pattern indicates that combinations of low (high) values of the dominant factor with high (low) values of its strongest interacting factor enhance PSS. This may be because these facilities tend to compete for space within the limited spatial confines of the 15-min living circle of ORAs [107]. For instance, both educational institutions and package stations intermittently require road access to facilitate pedestrian traffic and logistical operations. Superfluous parking provision may compromise medical service accessibility. Consequently, asymmetric density combinations of these facilities prove more appropriate within ORAs.

5.4. Retrofit Recommendations in ORAs

In the development of sentiment-friendly ORAs, from a spatial perspective, priority should be given to improving sentiment well-being in the four negative-sentiment agglomerations: the Tumen district, the north side district of the Daming Palace National Heritage Park, the Wanshou Road-Changlepo Metro Station contiguous area and the area east of Fangzhicheng, and the southern Ganjiazhai area. According to the analysis in Figure 9, the built environment impact pattern corresponding to these areas is Negative Sentiment—Insufficient Parking Facilities Dominant Type (Cluster 4). The renovation of this cluster of areas should focus on strengthening the construction of nearly all types of health facilities and balancing their interrelationships. Moreover, Figure 9 reveals that the environment-sentiment patterns in the 15-min living circle in ORAs exhibit inequitable characteristics. Priority should also be given to the renovation of Cluster 2 (Negative Sentiment—Insufficient Medical Facilities Dominant Type), where medical facilities, food facilities, and parking facilities should be strengthened. In addition, Cluster 3 (Positive Sentiment—Partial Health Facilities Promotion Type) areas are also key areas for improvement. Given that these areas cover a large area, renovation can effectively enhance overall sentiment. During the renovation process, efforts should be made to mitigate the impact of negative driving factors, such as park and plaza density.

In terms of facility development, it is recommended that infrastructure upgrades be prioritized in ORA retrofit projects. Subsequent interventions could focus on urban heat island mitigation, green space enhancement, and perception optimization in streetscape design. This is in accordance with the fact that in the process of promoting psychological well-being, meeting basic functional needs is more important than higher-order environmental and aesthetic demands. Emphasis should be placed on factors that have a significant impact on PSS. Fine-grained control over facility planning is essential, which allows for efficient and rational allocation of limited urban resources [109], avoiding both under-provision and resource wastage. Urban policymakers and planners are advised to shift focus from maximizing the quantity of facilities toward optimizing their density [115]. The objective is to identify and maintain an optimal density spot that best supports residents’ daily needs and promotes psychological well-being. Retrofitting must account for the nonlinear relationships between facility and sentiment. Interventions should be informed by threshold effects to maximize positive sentimental returns and avoid diminishing well-being benefits. Moreover, when deploying multiple facility types within the same neighborhood, planners should aim to exploit synergistic interactions while mitigating competitive interactions to help amplify positive sentimental outcomes and suppress negative externalities [116], ultimately contributing to health-promoting ORAs.

5.5. Limitations and Future Directions

Several limitations and further directions remain. First, Weibo check-in data is limited in obtaining users’ personal characteristics, such as gender, age, marital status, income, living arrangements, and residential duration [117]. This makes it difficult to compare the impact of built environment factors on sentiments with that of socioeconomic characteristics. Future research could explore the possibility of integrating external data sources that contain personal characteristics with Weibo check-in data.

Second, when using Weibo check-in data at the community level, the issue of spatially uneven data distribution is amplified compared to studies at the city level. This may affect the underrepresentation of specific locales with low social media penetration. We suggest that questionnaire data, positive and negative affective scale (PANAS), or self-reported sentiment data could be used or integrated in areas with low data density.

Third, certain factors were excluded due to significant differences in data granularity compared to the 100 m $\times$ 100 m grid used in this study, such as leaf area index, gross primary production, impervious surface area, and concentrations of PM_2.5, O₃, and NO₂, which are factors representing the physical environment in a community. Further research could explore larger geographical areas and coarser spatial resolutions to incorporate a broader range of factors, thereby achieving a more comprehensive understanding.

6. Conclusions

This study investigated the relationship between community health promotion factors and residents’ sentiments within the 15-min living circle of ORAs using BERT and XGBoost modeling.

We found: (1) Healthy facilities and old residential unit density emerged as the dominant influencing factors influencing sentiments in the 15-min living circle of ORAs. (2) Nonlinear relationships with threshold effects and saturation effects were observed in the effects of the dominant factors. Parking (threshold is 57.67), supermarket (threshold is 18.45), and cultural facility densities initially suppress and then enhance PSS as density increases with distinct thresholds. Sports facility density shows an initial promotive effect followed by suppression at higher densities. Education facilities and package stations densities demonstrate U-shaped relationships, while public transport station density and park and plaza density exhibit a four-phase pattern. (3) Interaction effects between facilities were observed in two patterns: promotion effects under co-presence conditions (parking with old residential unit, cultural facility with public toilet, public transport station with old residential unit, and park and plaza with public toilet promote PSS under low-low or high-high density combinations), and promotion effects under asymmetric conditions (supermarket with scenic spot, education with package station, and medical with parking enhance PSS under low-high or high-low density combinations). (4) Four distinct environment–sentiment impact types were identified: Positive Sentiment—Comprehensive Health Facilities Promotion Type, Positive Sentiment—Partial Health Facilities Promotion Type, Negative Sentiment—Insufficient Medical Facilities Dominant Type, and Negative Sentiment—Insufficient Parking Facilities Dominant Type. This classification provided decision-makers with a precise map and actionable guide for targeted interventions. These findings collectively advocate for a paradigm shift in ORA retrofitting: moving beyond quantity of dominant facilities toward optimizing density. This approach should consider the complex nonlinear relationships and threshold effects, leveraging the promotion effects of density. It should also enhance the synergistic effects between facilities while avoiding competitive interactions, ultimately improving livability and residents’ well-being.

Our study, focusing on ORAs, addresses the scale gap in existing research on the relationship between sentiment and the built environment, as well as in studies of healthy communities. The big data-based analysis eliminates biases associated with small samples and subjective evaluations, providing objective experimental results. The application of BERT, DeepLab V3+, XGBoost, and SHAP techniques realized parallel exploration of the health status of multiple ORAs from a macroscopic perspective. The comparative analysis of feature importance, nonlinear relationships, and interaction effects offers substantial evidence for the healthy renewal and provides valuable insights for the detailed development of the 15-min living circle of ORAs.