Measurement of spatial heterogeneity in street restorative perceptions and street refinement design

Abstract

Restorative perception of streets is an essential metric for evaluating urban environments and serves as a key indicator of pedestrians’ perspectives on street refinement design. However, restorative perception varies significantly across different streets, necessitating an analysis of these disparities. This study integrates street view data, deep learning algorithms, the MGWR model, and space syntax to analyze spatial heterogeneity in restorative perceptions and optimize street design strategies. Using the random forest (RF) algorithm and restorative component scales (RCS), we assessed residents’ perceptions of street restoration in Shanghai’s Huangpu District. Our analysis identified eight key visual elements influencing perceptions, such as sidewalks, buildings, and walls. Results revealed significant geographic variations, with high-perception areas concentrated around open parks, waterfronts, and well-designed buildings featuring thoughtful amenities. Visual elements like trees and plants were found to significantly enhance restorative perceptions. Moran’s Ι statistics and multiple regression models further revealed spatial heterogeneity and clustering in perceptions, highlighting the importance of location-based planning. Among the regression models, the MGWR model achieved the highest R² value (0.615), indicating that variables like trees, roads, sidewalks, and intercepts are particularly sensitive to spatial heterogeneity. Additionally, space syntax analysis underscores the positive impact of complex street networks on accessibility, convenience, and environmental satisfaction. The main contribution of this study is identifying the most effective model through a comparison of multiple regression models, demonstrating the spatial heterogeneity of different visual elements. Based on restorative perception and accessibility coupling assessment, streets in urgent need of rehabilitation were identified. We believe our findings can assist professionals in developing more targeted and effective strategies based on the restorative nature of streets.

Introduction

As a vital component of the urban landscape, streets serve not only as primary traffic routes but also as spaces for communication, shopping, walking, and social interactions^1,2. However, rapid urbanization has introduced challenges, including street sprawl, pollution, traffic congestion, and diminished street vitality³. Prolonged exposure to low-quality urban environments can lead to stress, which, if severe and prolonged, may cause mental health issues and physical illnesses⁴. Thus, urban planning efforts aimed at enhancing public mental health and well-being have become increasingly urgent. A key aspect of these efforts is understanding and improving restorative perceptions of the urban environment⁵. High levels of restorative perception in urban spaces are associated with reduced psychological stress, alleviation of negative emotions, and mitigation of chronic illnesses⁶.

The theoretical foundations of restorative perception in urban environments include the stress reduction theory (SRT) and the attention restoration theory (ART)⁷. Both theories emphasize the role of urban environments in helping individuals recover from attention fatigue and physiological-psychological stress⁸. SRT is a psychological and evolutionary model that focuses on stress caused by threats in urban environments. In contrast, ART centers on recovery from attention fatigue and is based on four core elements: effortlessness, fascination, coherence, and compatibility. ART also emphasizes how various elements, such as natural features and landscapes, support restorative perception. Based on these elements, Hartig et al. developed the perceived restorativeness scale (PRS)⁹, which is widely used in research to evaluate participants’ environmental restorative perception levels.

Early research on restorative environments primarily focused on urban green spaces, exploring their connections to human well-being, aesthetic preferences, and spatial features conducive to restoration. For instance, Konrad et al.¹⁰ assessed the restorative potential of soundscapes by asking 162 participants to listen to recordings from eight parks. Xue et al.¹¹ investigated the role of wellness tourism in shaping perceptions of restorative environments. Despite streets being among the most frequently used public spaces in urban life, their capacity to foster restorative perceptions has remained largely overlooked. In recent years, an expanding scientific perspective has prompted more extensive exploration of the restorative potential of urban streets. Scholars from various fields have examined how these essential urban spaces can support restoration. For example, Tabrizian et al.¹² proposed combining GIS high-resolution spatial data with photo-realistic immersive virtual environment perception studies to evaluate how street spatial configurations affect restorative perceptions. Meanwhile, Han et al.¹³ used virtual reality simulations and eye-tracking technology to demonstrate a positive link between street-side plants and restorative perceptions. Restorative perception is influenced by various factors, such as personal experiences and regional culture. Although these methods are accurate, they can be time-consuming, costly, and may produce limited results¹⁴. Furthermore, response bias complicates the analysis.

Recently, the development of crowdsourced mapping services and the availability of geotagged images from platforms such as Baidu Street View (BSV) and Google Street View (GSV) have introduced new methods for assessing urban street perceptions¹⁵. These images provide panoramic views of urban environments, serving as a low-cost, easily accessible, and comprehensive data source for assessing urban streets^16,17. For example, they can be used for large-scale analysis, helping to identify street features aligned with residents’ perceptions, thus deepening our understanding of urban environments¹⁸.

Additionally, with the rapid development of computer technology, some studies have systematically explored how to extract higher-level information from street view images¹⁹. First, deep learning has been used to extract semantic information from street images. Early methods relied on pixel colors to identify objects, but this approach had limitations when objects had similar colors or when rare objects appeared²⁰. In contrast, deep learning algorithms like FCN, ResNet, and SegNet utilize convolutional neural networks to process visual data, enabling the effective identification of various visual features such as lanes, buildings, skies, sidewalks, trees, and vegetation. This forms a strong foundation for research on urban street restorative and its relationship with human perception. Second, deep learning frameworks have been employed to assess restorative perception. Residents’ perceptions of urban street quality are based on their experiences. The human brain can quickly recognize complex scenes, providing a foundation for using deep learning models to conduct large-scale and rapid perception measurements²¹. For example, In the Place Pulse project at the MIT Media Lab, street view images were uploaded to a website (https://pulse.media.mit.edu), allowing visitors to compare images based on perceptual criteria such as safety, vitality, and wealth, generating a deep learning dataset for measuring street quality²². Wang et al.²³ proposed a deep learning-based human–computer adversarial framework using an RF module to explore the relationship between street view elements and user ratings.

The use of multiple linear regression models to explore the relationship between restorative perception and various explanatory variables has become a well-established method^24,25,26. However, many studies model cause-and-effect relationships in the global space. Street view image data with geographic coordinates serve as standardized spatial data²⁷, and studies using these images often reveal spatial heterogeneity. More importantly, global models based on such data may introduce errors and fail to accurately evaluate streetscape perceptions. In contrast, Geographically Weighted Regression (GWR) and Multiscale Geographically Weighted Regression (MGWR) models offer greater precision for spatial regression analysis by accounting for geospatial heterogeneity and applying local regression techniques²⁸. GWR captures spatial variations in the relationships between dependent and independent variables across locations²⁹. MGWR extends GWR by determining the optimal bandwidth for each explanatory variable, providing deeper insights into spatial relationships³⁰. Adopting appropriate spatial regression models has become a vital interdisciplinary research direction for spatial data analysis.

Furthermore, while integrating street-view images with deep learning is effective in assessing residents’ restorative perception of street spaces and provides valuable insights for urban planning, it is hindered by the lack of funding for urban renewal in many economically underdeveloped cities. This constraint limits broad renovations or reconstructions. Therefore, it is crucial to identify streets that urgently need updates to allocate funds more efficiently and enhance urban planning effectiveness.

This study aims to address three key objectives:

• 1. Utilize the RF algorithm, based on the RCS, to predict restorative perception scores of urban streets and identify the visual elements associated with different perception levels.

• 2. Compare the effects of street visual elements on restorative perceptions using the Ordinary Least Squares (OLS), GWR, and MGWR models to determine the most effective model and reveal spatial heterogeneity.

• 3. Examine the relationship between restorative perceptions and road accessibility using space syntax, analyze spatial heterogeneity, and provide refined urban design and governance recommendations.

Study area and data

Study area

Huangpu District, located in Shanghai, China, is a typical ultra-high-density urban space. The district covers 20.52 square kilometers, with 18.71 square kilometers being land. As of 2022, the residential land area is 6.42 square kilometers, and the permanent population is 662,030, resulting in a population density of 103,000 people per square kilometer (https://ghzyj.sh.gov.cn/). Huangpu is the historical and cultural heart of Shanghai, boasting the city’s richest heritage. It is also the only central activity zone in the city, with a unique core location. The district has a large economic volume, an optimal structure, and high density, yielding the most efficient economic output. These factors together make Huangpu a “dual-quality” space that blends Shanghai’s historical characteristics with modern urban qualities. As such, it serves as a key pilot area for urban renewal in Shanghai, offering practical application scenarios for our research and supporting relevant policy development. Most importantly, Huangpu provides diverse and easily accessible street-view image resources, which are valuable data for our research. These characteristics make the district an ideal location for using street-view data to assess how the urban street environment influences residents’ restorative perception while also emphasizing the vitality of street refinement in high-density, central urban areas(Fig. 1)³¹.

Fig. 1

Study area location: (A) China; (B) Shanghai; (C) Huangpu District; (D) Detailed map of Huangpu District. (Map created by author using Maps Data: Google, ©2025 Airbus, CNES/Airbus, Landsat/Copernicus, Maxar Technologies (https://earth.google.com/web/search/Shanghai/), and the street area was generated using GIS10.4.).

Data sources

Collecting baidu street view image data

Using ArcGIS 10.4, geographic collection points were generated at 50 m intervals along the street network of the study area. The street network data were sourced from OpenStreetMap (OSM) (https://www.openstreetmap.org/, accessed May 18, 2024). For each sample point, four photos (covering angles 0–90, 90–180, 180–270, and 270–360°) were obtained by extracting the coordinates and utilizing Baidu Street View (BAPI 3.0) through Python. The BSVI IDs for specific locations were retrieved using the following URL structure: (http://api.map.baidu.com/panorama/v2?ak=API_Key&width=600&height=400&location=LAT,LON&pitch=20&fov=90&heading=0/90/180/270). Here, LAT and LON denote the latitude and longitude, respectively. FOV specifies the horizontal field of view, HEADING indicates the compass direction of the camera, PITCH adjusts the vertical angle, and the API key authenticates the request. Finally, the program downloads street-view images from four angles for each coordinate point, capturing a total of 162,364 street-view images.

Restorative perception label data

To generate restorative perception labeling data, we randomly selected 500 sample images from the BSVI database that met the study criteria. Then, to ensure image quality and visual appeal, we excluded 8 images with poor weather conditions or low contrast²⁶. These images were presented to 30 participants, who were assessed using a Likert scale. Additionally, the selected images covered a wide range of environmental types to increase the generalizability of the study. We used the revised RCS as the primary tool for assessing street environments. The revised RCS consists of 15 questions related to the street environment (Table 1)^1,23. To measure restorative street perceptions, we developed a 5-level Likert scale questionnaire on the Wenjuanxing platform (https://www.wjx.cn/), covering four key dimensions: escape, extent, fascination, and compatibility^32,33. In the experimental phase, participants were asked to evaluate 500 randomly selected sample images online, with the instruction to observe each image for at least 10 s. Participants rated each image on a scale of 1 to 5 for each question, indicating how well their perceptions matched the images. They also provided demographic information, including age (18–28, 29–50, 50 +) and gender (Table 2). Finally, the ratings from all participants were aggregated into a labeled dataset, which was used to train the RF model. This approach to data labeling is more convenient, efficient, and cost-effective than traditional survey methods because it is conducted entirely online.

Table 1 The restorative perception scale.

Table 2 Descriptive statistics for the volunteers.

Urban street visual element data

This study employs semantic image segmentation techniques to calculate the percentage of semantic object elements in each street view image, aiming to analyze the relationship between specific locations and public perceptions of restoration through visual elements. A classical SegNet model, known for its effectiveness in scene segmentation, was trained using the ADE20K dataset (https://github.com/CSAILVision/ADE20K, accessed May 22, 2024), achieving an accuracy of 90.83% on the training set and 89.95% on the validation set.

In the SegNet configuration, images were resized to 416 × 416 pixels using the Reshape function³⁴. ZeroPadding2D was applied to the 2D image matrix, ensuring control over the feature map size and enabling efficient convolutional kernel extraction³⁵. Conv2D was used to extract features from the input high-dimensional arrays. BatchNormalization was employed to normalize the feature data, accelerating the gradient descent process, improving network training speed, and enhancing generalization³⁶. MaxPooling2D was used to reduce the image dimensions and the number of neuron parameters while preserving the key features of the image. UpSampling2D restored the image to its original size while the encoder resized it to 208 × 208 pixels. Softmax was applied to compute the semantic classification probability for each pixel across 150 categories. The code for these experiments is available on GitHub. (https://github.com/CSAILVision/semantic-segmentation-pytorch, accessed on January 22, 2024). An overview of the semantic segmentation of the street view images is presented in (Fig. 2).

Fig. 2

An overview of semantic image segmentation utilizing SegNet trained on ADE20K dataset: (A) input BSVI; (B) SegNet framework; (C) output image featuring the segmentation mask.

Methodology

First, an RF-based machine learning algorithm was utilized to assess urban street restorative perception scores. Second, OLS regression analysis was applied to evaluate whether restorative perception scores and visual elements exhibited spatial heterogeneity. Subsequently, GWR and MGWR regression analyses were conducted to identify the model with the highest explanatory power and explore the relationship between street visual elements and restorative perception. Third, spatial heterogeneity was further examined by analyzing the relationship between restorative perception and street accessibility. Finally, refined design and governance recommendations were proposed. An overview of the methodology is illustrated in (Fig. 3).

Fig. 3

The research flowchart.

The RF-based restorative perception prediction of urban street design

We employed an RF-based deep learning algorithm to investigate residents’ evaluations of restorative perception in urban streets. Each volunteer rated the restorative perception of 500 street view images, and the aggregated ratings from all volunteers were used as the training dataset for the RF model. During training, the bootstrap method randomly selected two-thirds of the samples for data fitting or classification. In contrast, the remaining one-third, defined as out-of-bag (OOB) data, was used to evaluate the overall model error and the importance of the variables³⁰. Suppose the input variable ${X}_{j}$ is entered into the equation. In that case, to calculate the significance of the input variable ${X}_{j}$ in the ${\text{Nth}}$ tree ${VI}_{n}$, it is necessary to create the regression tree model ${T}_{n}$ using the sample data extracted by Bootstrap and calculating the prediction error for OOB again before randomly substituting the observations in the variable $Xj$. Finally, the model ${T}_{n}$ is rebuilt to calculate the prediction error of OOB. By processing the prediction errors of the two OOB data, the average of all regression tree results is the critical variable $Xj$ in the ${\text{Nth}}$ random tree ${VI}_{n}\left({X}_{j}\right)$. It is calculated according to Eq. (1) ref.³⁰.

$$V{I}_{n}\left({X}_{j}\right)=\left\{\sum_{i=1}^{{N}_{OOB}}I\left[f\left({X}_{i}\right)={f}_{n}\left({X}_{i}\right)\right]-{\sum }_{i=0}^{{N}_{OOB}}I\left[f\left({X}_{i}\right)={f}_{n}\left({X}_{i}{\prime}\right)\right]\right\}/{N}_{OOB}$$

(1)

Urban street restorative interpretation model

To investigate the influence of urban street visual elements on restorative perception and their spatial distribution, we applied three regression models: OLS, GWR, and MGWR. The model with the highest explanatory power was selected for further analysis.

OLS is widely employed as a linear regression method to examine the relationship between dependent and independent variables. As a global model, OLS minimizes the sum of squared differences between observed and predicted values in the dataset to estimate the parameters of explanatory variables³¹. The model assumes spatial stationarity, indicating that data points’ geographic position or spatial distance does not influence the regression outcomes³². However, in practical applications, many explanatory variables exhibit spatial non-stationarity, meaning their explanatory power varies across different geographic locations. If the Koenker (BP) test indicates statistical significance (p < 0.05), it implies that the data exhibits spatial non-stationarity, suggesting the necessity of using a spatial regression model for analysis³³. The GWR model is designed to effectively address spatial non-stationarity in regression analysis³⁵. The MGWR model assigns distinct bandwidths to each explanatory variable, allowing for adjustments across spatial scales and thereby reducing parameter estimation errors³⁴.

First, this study identifies the top eight street visual elements ranked by their influence on restorative perception from the image segmentation data and uses them as explanatory variables. Subsequently, OLS regression is applied for global analysis to examine the relationships among the variables. The formula is as follows¹.

$${Y}_{i}={\alpha }_{0}+\sum_{k=1}^{m}{a}_{k}{X}_{k}+{\varepsilon }_{i}$$

(2)

where ${Y}_{i}$ represents the restorative perceptual score and its coordinates $\left({u}_{i},{v}_{i}\right)$; ${\alpha }_{0}\left({u}_{i},{v}_{i}\right)$ indicates the model’s intercept; ${a}_{k}$ represents the regression coefficient of the $kth$ independent variable data; ${X}_{k}$ is the $kth$ attribute; ${\varepsilon }_{i}$ is the random error of the model.

Exploring explanatory variables with OLS models can help determine whether they exhibit spatial smoothness. If they do not, this indicates spatial heterogeneity and necessitates regression analysis using a spatial model. The formula for GWR is provided below²⁰:

$${Y}_{i}={\beta }_{0}({u}_{i},{v}_{i})+\sum_{k}{\beta }_{k}({u}_{i},{v}_{i}){X}_{ik}+{\varepsilon }_{i}$$

(3)

where ${Y}_{i}$ represents the dependent variable, an independent variable, $\left({u}_{i},{v}_{i}\right)$ is the geographic coordinate of the $ith$ point, ${\beta }_{0}\left({u}_{i},{v}_{i}\right)$ is the intercept of the $ith$ point, ${\beta }_{k}$ $\left({u}_{i},{v}_{i}\right)$ is the coefficient of ${X}_{ik}$, and ${\varepsilon }_{i}$ is the residual of the $i$ point.

The GWR model allows for observing variations in regression coefficients for restorative perception across different spatial locations³⁷. These variations provide insights into how explanatory variables influence the restorative perception of urban streets in diverse spatial contexts. Defining the neighborhood range, which specifies the data points included within a given scope, is crucial for applying the GWR model. The model uses a single fixed bandwidth to delineate this range, assuming that the regression coefficients for all explanatory variables change consistently across the same spatial scale³⁸. The bandwidth, measured in metric units of distance, utilizes a bi-square kernel to assign weights to data points within its range, with weights determined by the inverse distance from the center of the local region. Data points outside the bandwidth are assigned a weight of zero. A golden section search optimization algorithm is used for parameter optimization, with the AIC serving as the model selection criterion. The model parameters are evaluated using local R² and global AIC values³⁹. Although GWR employs a fixed bandwidth to estimate regression coefficients, the explanatory variables may exhibit spatial variability.

The MGWR model allows each explanatory variable to use an independent bandwidth, with the bandwidth size adapted to its respective spatial scale. By accounting for the varying spatial scales of influence associated with each explanatory variable, this model reduces errors in parameter estimation¹. The MGWR formula is expressed as:

$${Y}_{i}={\beta }_{0}\left({u}_{i},{v}_{i}\right)+{\sum }_{k=1}^{m}{\beta }_{bwk}\left({u}_{i},{v}_{i}\right){X}_{ik}+{\varepsilon }_{i}$$

(4)

where $\left({u}_{i},{v}_{i}\right)$ denotes the geographic coordinates of the $ith$ point; ${\beta }_{bwk}$ denotes the regression coefficient of the explanatory variable $k$; m is the number of sampling points; ${\beta }_{bwk}$ is the bandwidth; and ε_i is the residual of the $i\text{th}$ point.

Finally, we compared the explanatory power and AIC values of the OLS, GWR, and MGWR models to determine the most suitable model for analyzing restorative perception and street view regression. This comparison ensures that the selected model effectively captures spatial variations, providing reliable insights into how urban street visual elements influence restorative perceptions.

Spatial syntax-based street accessibility analysis

This study applies space syntax to design complex street views as mathematical graphs with multiple nodes to investigate their spatial structural features. Two models are commonly used in spatial analysis of urban spaces: the axial line model and the segment model. Previous studies have shown that residents’ activity radii typically fall within a 500 m range⁴⁰. Therefore, we used DepthmapX, setting Choice to 500 m, to calculate the accessibility of streets within a 500-m radius in the study area. However, the complexity of the road data may lead to inaccuracies in accessibility calculations. To address this, we use GIS 10.4 to buffer roads, extract centerlines, and construct a simplified road network⁴¹. The streets are then merged, simplified, and topologically processed to generate road centerlines with accurate topological structures, providing reliable preliminary data for accessibility calculations.

Results

RF algorithm accuracy testing

To ensure the robustness of the RF model and mitigate the risk of overfitting, we adopted two independent validation methods. First, we randomly divided the labeled dataset into a training set (two-thirds of the samples) and a test set (one-third of the samples)^19,23. The test set was used to evaluate the model’s performance on unseen data. Second, we utilized the Out-of-Bag (OOB) error, an internal cross-validation method in the RF algorithm, to assess overall prediction stability. The results show that the model achieves a Root Mean Square Error (RMSE) of 1.15 on the test set. In addition, the OOB error is 2.79%, indicating good generalization ability. To provide a more comprehensive evaluation, we also calculated the Mean Absolute Error (MAE) and the coefficient of determination (R²). The RF model achieved an MAE of 0.94 and an R² of 0.82 on the test set, demonstrating its effectiveness in predicting residents’ restorative perceptions.

Spatial distribution of street restorative perceptions

High restorative perceptions are associated with high-quality urban streets, while low restorative perceptions indicate a need for focused attention. Therefore, we categorized the perceptual scores predicted by the RF model into three levels: scores below one as low restorative perception streets, scores between two and three as medium restorative perception streets, and scores above four as high restorative perception streets. Figure 4 displays the streets categorized by high, medium, and low restorative perception levels. Figure 5 illustrates the spatial distribution of restorative perception.

Fig. 4

Examples of Street View Images representing different high, medium, and low scores.

Fig. 5

The spatial distribution of restorative perception. (Map created by author using GIS10.4 and Mapbox: https://www.mapbox.com/mapbox-studio) .

The data were clustered into seven classes to further examine the geospatial distribution of high and low restorative perceptions of streets in the study area. As shown in Fig. 6, there is a clear geospatial difference between areas with high and low restorative perceptions. Areas A and D were identified as regions with negative restorative perceptions. Area A primarily includes the longitudinal Middle Sichuan Road, Huqiu Road, Middle Jiangxi Road, and the intersecting Xianggang Road. Area D is centered around Huaxingli, with several alleyways, such as Caoshi Alley, Menghua Street, and Wangyun Road, identified as spaces with negative perceptions. These areas are characterized by informal street construction, littering, tangled overhead cables, lack of pedestrian walkways, traffic congestion, and dilapidated buildings. Area B represents a positively perceived space, forming a rectangular area of 1.26 km² around People’s Park, Yanzhong Greenbelt, and Huaihai Park. It is bounded by Jiujiang Road to the north, Middle Xizang Road to the east, Middle Huaihai Road to the south, and North Huangpi Road to the west. The area’s proximity to parks, municipal institutions, and a predominantly commercial environment contributes to a more livable environment and better infrastructure. Area C, also identified as a positive perception space, is centered around Zhongshan No. 1 Road. To the east, it is bordered by a 50-m-wide waterfront promenade along the Huangpu River, providing urban residents with rich recreational space and a natural environment. To the west, it is adjacent to the Bund architectural complex, a stunning collection of historical buildings that blend classical and modernist styles, which enhances its positive perception. Area E encompasses a waterfront space between Miaojiang Road in the east and Jiangbin Road in the west, connecting open parks such as the South Park Riverside Green Space, Huangpu Riverside Park, and the Chinese National Rejuvenation Theme Park. Large exhibition complexes, including the Children’s Art Center, the Shanghai Museum of Contemporary Art, and the World Expo Museum, are located near these open park spaces. These cultural landmarks contribute to an optimistic residents’ perception.

Fig. 6

The restorative perception clustering with different scores. (Map created by author using GIS10.4 and Mapbox: https://www.mapbox.com/mapbox-studio).

Table 3 presents the eight street view elements with the highest proportions identified during image segmentation. These elements are considered most likely to influence restorative perception positively. The three elements with the highest proportions are ‘sidewalk’, ‘building’, and ‘wall’, while those with the lowest are ‘road’, ‘tree’, and the ‘sky’.

Table 3 The top eight visual elements were identified after the segmentation of BSVIs. To visualize the results clearly, we chose to keep three valid digits for our study data.

We averaged the street visual elements with high, medium, and low restorative perception scores to create a per-street visual element percentage feature for each perception category (Fig. 7). Our analysis revealed that ‘ceiling’ and ‘building’ have a relatively low proportion of high restorative perception ratings. We hypothesize that viaducts and high-density buildings may create a sense of oppression by disrupting the original city skyline, reducing natural light, narrowing the field of view, and making the space feel small and enclosed, particularly in older urban areas. These abrupt modern structures may clash with the scale and style of historical buildings, exacerbating the psychological sense of confinement and resulting in lower restorative perceptions. In contrast, elements such as ‘tree’, ‘plant’, and ‘sidewalk’ are more frequently associated with high restorative perceptions. ‘Tree’ creates a more comfortable and natural environment by improving visual appeal, air quality, and noise control, which enhances mood and reduces fatigue. A higher presence of plants often signifies greater spatial openness, promoting a sense of freedom and reducing stress and anxiety, thereby improving emotional well-being. Additionally, well-designed sidewalks enhance the walking experience by balancing comfort, safety, and aesthetics with the surrounding environment, fostering a pleasant, relaxing atmosphere that increases restorative perceptions.

Fig. 7

Percentage of visual elements with low, medium, and high restorative perceptions of the street.

Spatial heterogeneity in street restorative perceptions and its influencing factors

This study initially employed the spatial autocorrelation function in ArcGIS 10.4 to perform a global analysis of spatial patterns in restorative perception. As shown in Table 4, Moran’s l with P below 0.01 indicates significant spatial autocorrelation in restorative perceptions⁴², reflecting a 99% confidence level for the observed spatial clustering.

Table 4 Spatial autocorrelation Moran I statistical table.

Subsequently, we applied local spatial autocorrelation to further analyze the presence of spatial autocorrelation in street restorative perception. This analysis classified local spatial autocorrelation into four types: low-low clustering, high-low clustering, low–high clustering, and high-high clustering³⁹. Figure 8 illustrates how these clusters relate individual streetscape viewpoints to surrounding spatial perception types, offering a clearer and more intuitive visualization of the spatial autocorrelation distribution in urban street perceptions³⁰. Specifically, “high-high” clustering indicates that areas with high perception values are surrounded by other high-perception areas, reflecting a strong spatial aggregation effect. These “high-high” clusters are predominantly found along the Huangpu River and near open green spaces. Conversely, “low-low” clusters—where low-perception points are adjacent to other low-perception points—also exhibit spatial aggregation effects. These “low-low” clusters are mainly concentrated in the Laoximen sub-district. Additionally, “high-low” and “low–high” clusters suggest no significant spatial proximity effects.

Fig. 8

Spatial autocorrelation clustering. (Map created by author using GIS10.4 and Mapbox: https://www.mapbox.com/mapbox-studio).

Table 5 indicates that OLS regression can determine whether there is spatial heterogeneity in the data, serving as a preliminary check for spatial regression. A VIF < 7.5 suggests that there are no covariance issues. Therefore, we excluded the ‘building’ due to spatial covariance problems.

Table 5 Summary of OLS results (model variables).

Table 6 indicates the results of the Koenker (BP) test, which is significant. The significance of each visual element was assessed using Robust_Pr, with P < 0.05. This means that ‘wall’ and ‘plant’ should be excluded, while the six street visual elements—‘intercept’, ‘road’, ‘sky’, ‘tree’, ‘sidewalk’, and ‘ceiling’—are statistically significant. The significance of the Koenker (BP) test results demonstrates that the different street visual elements are spatially non-collinear, indicating that the same variables have varying fitting effects across different spatial regions, thus exhibiting heterogeneity. Consequently, GWR and MGWR models are appropriate for the subsequent spatial regression analysis.

Table 6 OLS diagnostics.

Table 7 shows that, compared to the OLS model, both GWR and MGWR significantly improve the model’s goodness of fit. The R² value increases from 0.378 in OLS to 0.566 in GWR and 0.615 in MGWR, reflecting a progressive enhancement in the model’s explanatory power. Furthermore, the AIC value decreases substantially from 12,171.752 in OLS to 9,099.070 in GWR and further drops to 8,913.321 in MGWR. The lowest AIC value in the MGWR model indicates an optimal balance between model complexity and goodness of fit, confirming its superior performance. Notably, regardless of the model used, ‘road’, ‘sky’, ‘tree’, ‘sidewalk’, and ‘ceiling’ consistently exhibit a positive correlation with restorative perception, indicating that these visual elements contribute positively to restorative perception.

Table 7 Model regression results.

We used violin plots to further analyze the MGWR model’s regression coefficients (Fig. 9). The ‘intercept’ coefficient’s highest density region lies below zero, indicating that the intercept negatively affects restorative perception in most areas, with positive perceptions occurring only in a few spatial regions. The highest density regions of coefficients for ‘road’, ‘sky’, ‘tree’, ‘sidewalk’, and ‘ceiling’ are all greater than zero, suggesting that these coefficients positively influence restorative perception in most areas. Notably, the highest density region of the coefficient distribution for trees is ≥ 0.5, highlighting the significant impact of trees on restorative perception.

Fig. 9

Violin plot of MGWR model regression coefficient distribution.

In addition, geographically, ‘tree’, ‘road’, ‘sidewalk’, and ‘intercept’ exhibit greater spatial heterogeneity than other elements, as reflected by the coefficients in the MGWR model (Fig. 10). The placement of trees near Wusheng Road and South Huangpi Road significantly enhances restorative perception. At the intersection of the North–South Elevated Road and Hefei Road, road quality strongly influences residents’ restorative perception. Improving viaduct spaces through landscape design and increasing street greenery can alleviate feelings of oppression and enhance overall comfort. ‘Sidewalks’ play a crucial role in shaping perceptions around People’s Square, where various streetscape elements and walkable sidewalks contribute to positive perceptions. Conversely, the ‘intercept’ has a significantly negative impact on perceptions in areas such as the circular road network formed by Hubei Road, Middle Xizang Road, and Haikou Road, the intersection of Yan’an Elevated Road and North–South Elevated Road, and the Laoximen sub-district. Strengthening infrastructure in these regions, by enhancing public service facilities and increasing green spaces, is recommended. On the other hand, the ‘ceiling’ and ‘sky’ have relatively weak spatial heterogeneity in their impact.

Fig. 10

The spatial distribution of the restorative perception’s influencing factors. (Map created by author using GIS10.4 and Mapbox: https://www.mapbox.com/mapbox-studio).

The relationship between street restorative perception and accessibility

Figure 11 illustrates the accessibility of areas within a 500 m radius of residents’ daily walking routes. Major streets with high complexity and density are shown as high-accessibility areas (marked in red), while streets with lower accessibility are indicated in green²³. High-accessibility areas are mainly concentrated in the northeastern corner of the study area, between Middle Xizang Road and Middle Henan Road, as well as within the circular area formed by Renmin Road and Zhonghua Road.

Fig. 11

The spatial distribution of the street accessibility within walking distance of 500 m. (Map created by author using GIS10.4, Depthmap-Beta1.0, and Mapbox: https://www.mapbox.com/mapbox-studio).

To evaluate the quality of the urban environment by integrating street accessibility and restorative perception and to promote refined urban design. We overlaid the top 20% of areas with high accessibility, the top 20% (high spatial perception), and the bottom 20% (low spatial perception) of spatial perception⁴¹. As shown in Fig. 12, this analysis highlights two key areas: high accessibility with high perception and high accessibility with low perception. The former is mainly concentrated in Districts B and D, while the latter is primarily found in Districts A and C. This pattern aligns with the overall distribution of street restorative perception in Huangpu District. Areas with both high accessibility and perception (areas B and D) feature high-quality commercial buildings, municipal institutions, open parks, and well-developed public infrastructure.

Fig. 12

The spatial distribution of the perception performance with a high accessibility. (Map created by author using GIS10.4, Depthmap-Beta1.0, and Mapbox: https://www.mapbox.com/mapbox-studio).

In contrast, areas with high accessibility but low perception (areas A and C), which are among the earliest downtown areas in Shanghai, are characterized by unauthorized street construction, mixed-use overhead cables, lack of sidewalks, deteriorating buildings, and aging infrastructure. Additionally, the layout of collective buildings and pedestrian spaces is poorly planned. Consequently, these high-accessibility, low-perception areas require focused urban renewal and design interventions.

Discussion

Differences in the geospatial distribution of high and low restorative perceptions in Huangpu district

In examining the distribution of restorative perceptions of urban streets in the Huangpu District, we found significant geographic variations between areas with high and low restorative perceptions. These differences are strongly associated with urban elements such as buildings, sidewalks, public spaces, and infrastructure. Highly restorative perceptions are concentrated near open parks, waterfront spaces, commercial areas with abundant recreational facilities, and high-quality buildings. Specifically, in Area B, high restorative perceptions are centered around open parks such as People’s Park, Yanzhong Greenbelt, Huaihai Park, and Taipingqiao Park, which are surrounded by a rectangular road network covering 1.26 km². These greenbelts and public spaces provide a pleasant natural environment for residents, offering places for leisure, shopping, and socializing. Area C is bordered by a 50-m-wide waterfront landscape belt along the Huangpu River to the east and the historic architectural cluster of the Bund to the west. In this area, residents experience the prosperity and multicultural vibrancy of Shanghai as an international metropolis from the early twentieth century while engaging in leisure activities along the open waterfront and enjoying views of the Lujiazui International Financial Center, embodying the modern city’s dynamism and technological advancements. In Area E, a waterfront space along the Huangpu River connects several open green spaces, including Nanyuan Riverside Park, Huangpu Riverside Park, and the Chinese Nation Rejuvenation Park. Additionally, modern exhibition-oriented buildings, such as the Children’s Art Center, the Shanghai Museum of Contemporary Art, and the World Expo Museum, are planned along nearby Miaojiang Road and Jiangbin Road. This layout caters to residents’ entertainment, leisure, and learning needs, providing a fulfilling perceptual experience.

In contrast, restorative perception scores were lower in Areas A and D. Area A primarily consists of north–south oriented roads, such as Middle Sichuan Road, Huqiu Road, Middle Jiangxi Road, and Xianggang Road, which intersect with east–west roads. Area D is centered around Huaxingli Lane and extends throughout the surrounding road network, including Caoshi Alley, Menghua Street, and Wangyun Road. Both Areas A and D were once cultural and commercial centers of Shanghai, with histories dating back to the late 19th and early twentieth centuries. However, with the city’s modernization, especially following the economic reforms of the 1990s, urban development gradually expanded outward, and substantial resources and investments were directed toward developing new districts. Due to historical challenges, these areas now face high population density, limited land resources, and a lack of effective urban design updates, resulting in outdated infrastructure. For instance, garbage piles, informal structures along the streets, damaged roads, and a lack of public service facilities contribute to residents’ depressing and noisy environment.

By analyzing the clustering of restorative perceptions in the study area, we found that variations in the geospatial distribution of high and low restorative perceptions reveal disparities in the availability of essential public facilities and services across different areas. Areas with high restorative perceptions are generally characterized by well-developed infrastructure and high-quality public services, while those with low restorative perceptions often lack these essential urban elements. These clustering patterns underscore the pivotal role of urban design and management in enhancing residents’ restorative perceptions. Effective urban strategies can enhance resource distribution, upgrade infrastructure, and improve residents’ quality of life and spatial perceptions. Furthermore, research has demonstrated that ‘tree’ and ‘plant’ contribute to residents’ subjective well-being by significantly reducing the urban heat island effect, minimizing noise, and improving air quality—key determinants of spatial perception2. Similarly, a well-designed ‘sidewalk’ can dramatically boost restorative perceptions by providing a more enjoyable walking experience, combining comfort, safety, and aesthetics in harmony with the surrounding environment.

Spatial autocorrelation analysis of Huangpu district and its impact on improving urban restorative perception

We conducted a spatial autocorrelation analysis using Moran’s I, revealing a significant clustering pattern in restorative perception across the study area. This indicates that residents’ perceptions are influenced by the surrounding urban environment rather than being randomly distributed. The observed spatial dependence highlights the significant influence of the urban environment on residents’ psychological well-being.

Local spatial autocorrelation analyses show that high-high clusters are primarily located in areas with superior infrastructure and appealing environments, while low-low clusters are concentrated in regions with inadequate infrastructure and poor environmental conditions. These patterns suggest that residents’ spatial perceptions are shaped by both their immediate surroundings and the broader ecological context. High-low clusters often arise when high-quality facilities are situated in areas with initially low perceptions, potentially enhancing local perceptions. Conversely, low–high clusters may occur due to insufficient regional infrastructure, leading to a decline in overall perceived quality.

Street visual elements have a significant impact on spatial restorative perception, with components such as ‘building’, ‘road’, ‘sidewalk’, and ‘intercept’ playing crucial roles. High-quality ‘building’, ‘road’, and ‘sidewalk’ contribute to aesthetic appeal and a sense of security. In contrast, ‘intercept’ negatively impacts restorative perception, likely due to outdated infrastructure in older urban areas, including overhead cables, unauthorized street constructions, inadequate transportation systems, and deteriorating buildings. These elements can cause visual and psychological discomfort, fostering feelings of oppression and anxiety. These findings underscore the necessity of urban renewal and effective management to improve restorative perceptions. Urban planners can enhance residents’ quality of life by strategically allocating resources, optimizing infrastructure, and upgrading public services. Targeting specific areas for improvement can foster environments that support positive restorative perceptions, alleviate psychological stress, and enhance overall well-being. The spatial autocorrelation analysis further suggests that constructing more open green spaces, upgrading infrastructure, and enhancing public services are essential for improving residents’ restorative perception of street environments.

The effect of street visual elements and accessibility on spatial heterogeneity in Huangpu District

The analysis of the different regression models revealed that the R² value for MGWR was 0.615, compared to 0.566 for GWR and 0.378 for OLS. This indicates that the MGWR model is more suitable for exploring the relationship between restorative urban street perceptions and visual elements. The MGWR regression results highlight that the ‘tree’, ‘road’, ‘sidewalk’, and ‘intercept’ exhibit greater sensitivity to spatial heterogeneity than other visual elements. These spatial variations can be attributed to the differing demands and sensitivities of various regions to these elements. For example, the tree near Wusheng Road and South Huangpi Road demonstrates higher sensitivity, significantly enhancing residents’ visual and psychological experience. Increasing tree coverage in these areas could lead to a substantial improvement in restorative perceptions.

‘Road’ exhibits high sensitivity at the North–South Elevated Road and Hefei Road intersection. Enhancing spatial design and visual elements beneath the elevated roads through innovative landscape design can transform the area from a monotonous transit corridor into a vibrant community asset. This reimagining of elevated roads—from “barriers” to “resources”—provides residents with opportunities for relaxation and fosters a sense of belonging in the bustling urban environment. Such improvements enhance the overall quality of life and increase residents’ satisfaction with their surroundings. Urban renewal and future design efforts should prioritize constructing roads in these regions to provide residents with an improved commuting and travel experience.

‘Sidewalk’ near People’s Square has a significant impact on restorative perceptions. Features such as diverse vegetation, robust safety standards, and well-maintained infrastructure enhance residents’ restorative experiences. Pedestrian-friendly walkways transcend their role as mere passageways, evolving into public spaces that seamlessly combine comfort, safety, aesthetics, and functionality. These designs not only meet practical needs but also enhance the walking experience, fostering a stronger sense of community, belonging, and satisfaction.

‘Intercept’ negatively impacts residents’ restorative perceptions, particularly in areas like the circular road network formed by Hubei Road, Middle Xizang Road, and Haikou Road; the intersection of Yan’an Elevated Road and the North–South Elevated Road; and the Laoximen sub-district. Outdated infrastructure can inconvenience residents daily, leading to safety hazards, environmental degradation, and limited economic growth¹. Upgrading infrastructure in these areas—such as improving road accessibility, repairing dilapidated buildings, and adding lighting, green spaces, and public facilities—can create functional, visually appealing, and vibrant walking environments, thereby boosting residents’ restorative perceptions.

In contrast, the spatial heterogeneity of the ‘ceiling’ and ‘sky’ has a relatively weak influence but still still contributes to residents’ restorative perceptions. Factors such as sky openness, the landscape design of viaduct roads, soundproofing measures, safety, and air quality influence residents’ restorative perceptions. Urban planners should consider these factors holistically, designing more dynamic and appealing urban spaces by incorporating varied street layouts and strategically positioning key elements.

High accessibility areas are primarily located between Middle Xizang Road and Middle Henan Road and within the circular road network formed by Renmin Road and Zhonghua Road, according to the accessibility analysis of urban streets. Dense and intricate street networks provide residents with convenient travel options and various lifestyle services. High accessibility enhances residents’ convenience and their perception of environmental restoration. Residents in high-accessibility areas can efficiently access numerous services and facilities, significantly boosting their environmental satisfaction and spatial awareness.

Integrating urban street accessibility and restorative perception analysis reveals that areas with high accessibility and high restorative perception are primarily situated near open park spaces, waterfront areas, cultural institutions, and high-quality buildings. Conversely, areas with high accessibility but low restorative perception are concentrated in the earliest urban centers, such as around Huaxing Lane in the Laoximen sub-district. Urban renewal and governance should prioritize these areas. Enhancing infrastructure and refining spatial layout can improve the overall perceived quality of these areas. Research indicates that residents in accessible, well-served areas report significantly higher life satisfaction and well-being, highlighting the importance of improving street accessibility to boost restorative perceptions.

Conclusion

This study aims to examine the relationship between restorative perceptions of urban streets and various visual elements in the street environment to enhance residents’ restorative perceptions and foster harmony between residents and the city. First, using large-scale urban streetscape data and deep learning algorithms, we developed a framework for measuring restorative perceptions of urban streets, addressing the challenges of assessing perceptions in complex, large-scale urban environments. Second, the MGWR model was employed to analyze how the street environment affects residents’ restorative perceptions and to evaluate the impact of spatial heterogeneity. Third, street accessibility and restorative perception were analyzed using space syntax to investigate strategies for refining urban design. This study fosters collaboration between urban planners and mental health researchers. The proposed framework enables researchers to efficiently collect extensive data, measure and map restorative perceptions of urban environments, and develop strategies to enhance residents’ restorative perceptions from interdisciplinary and practical perspectives.

The findings show that the quality of public spaces, buildings, sidewalks, and infrastructure significantly influences residents’ spatial perceptions. Areas with high restorative perceptions are typically characterized by open green spaces, abundant recreational amenities, well-developed infrastructure, and high-quality buildings. For instance, Area B features a rectangular road network near People’s Park, Yanzhong Green, Huaihai Park, and Taipingqiao Park. Area C, centered on East Zhongshan Road No.1, is flanked by the Huangpu River waterfront to the east and the Bund’s historical architectural complex to the west. Area E includes several modern exhibition-oriented buildings along Miaojiang Road and Riverside Road. Conversely, areas with low restorative perceptions are concentrated in traditional neighborhoods in the old city center, characterized by dilapidated buildings and outdated infrastructure, such as Caoshi Alley, Menghua Street, and Wangyun Road near Area D.

Spatial statistical analysis with Moran’s I identified significant clustering in restorative perceptions, emphasizing the impact of the urban environment and underscoring the need for targeted urban design. High restorative perception areas are predominantly located in regions with well-developed infrastructure and abundant greenery. In contrast, low restorative perception areas are strongly linked to inadequate infrastructure and poor environmental conditions.

By Further analyzing the specific effects of various streetscape visual elements on residents’ restorative perceptions, we find that planting trees in appropriate areas significantly enhances residents’ sense of aesthetics and comfort. Well-designed roads and high-quality buildings enhance perceptions of safety and vitality. Conversely, outdated infrastructure, including limited public spaces, informal street constructions, deteriorating facades, and congested roads, negatively impacts perceptions, fostering feelings of oppression and boredom. MGWR analysis highlights the spatial heterogeneity in how street visual elements impact perceptions, offering valuable insights for targeted urban design.

The study also revealed that well-connected areas with intricate, dense street networks improve residents’ accessibility and service availability, leading to higher restorative perceptions. In contrast, less accessible regions should prioritize urban planning and governance to enhance infrastructure, spatial design, and layout.

Building on these findings, we present the following specific recommendations:

• 1. Expand open green spaces and public areas, including neighborhood parks, waterfront parks, and wetland parks. Enhancing greenery and accessible leisure spaces can foster a harmonious interaction between people and nature, improving visual appeal and psychological well-being.

• 2. Improve road quality and pedestrian walkability. Enhance traffic infrastructure for smoother transit while enriching streetscapes along sidewalks. Open roadside green spaces and create warm, comfortable, peaceful public spaces.

• 3. Promote thoughtful architectural design and street space design. Utilize well-planned designs and layouts to improve safety and aesthetic appeal.

• 4. Minimize negative visual elements. Focus on upgrading infrastructure in high-density residential areas, improving public services, living conditions, transportation, and municipal facilities while fostering economic and commercial development.

This study highlights the potential of integrating streetscape data, deep learning algorithms, MGWR models, and space syntax to assess urban restorative perceptions and guide refined urban design. By examining the spatial distribution and key factors influencing restorative perceptions, urban planners can design targeted strategies to create livable, resilient urban spaces, improving restorative perceptions and supporting sustainable development.

Although this study provides valuable insights, some limitations remain. First, two-thirds of the restorative perception scores were used to train an RF model, following the approach of Yao et al.^19,23, which is widely cited in urban planning. They invited 20 volunteers, each of whom was required to evaluate over 1000–2000 street-view images to train the model. However, in our study, it was challenging for the invited volunteers to focus on evaluating more than 500 images, and as the number of evaluations increased, the arbitrariness of their assessments also grew, leading to unstable model accuracy. Since this study used a small sample for training, model instability, and generalizability bias may have occurred. In the future, with sufficient funding, we plan to expand the study to cover central Shanghai, collect more street view data, and increase the sample size for model training. Second, this study did not fully consider the potential impact of seasonal changes on people’s perceptions, especially changes in plant life, when collecting streetscape images. Third, although we invited participants of different genders, ages, and occupations, we did not analyze the differences in restorative perceptions among different demographic groups.