Spatial Mismatch Between Street Space Quality and Public Sentiment: A Multilevel Spatial Diagnostic Approach

Abstract

Against the background of urban renewal, urban governance has increasingly pursued the coordinated development of spatial provision and public needs. Taking the central urban area of Shanghai as a case study, this study integrates street-view imagery and geotagged Weibo data to examine the spatial matching relationship between Street Space Quality (SSQ) and public sentiment. The results show that although SSQ and public sentiment both exhibit significant spatial clustering, no stable synchronization is observed between their spatial patterns. Most mismatches remain localized, whereas some display clear neighborhood continuity; moreover, the stable matching types exhibit distinct SSQ profiles. Overall, spatial mismatch in central Shanghai is characterized by the coexistence of locality and neighborhood continuity and by pronounced differences in quality profiles across types. These findings provide a spatial analytical basis for coordinated diagnosis and fine-grained governance in urban renewal.

1. Introduction

As urbanization shifts from incremental expansion toward the optimization of existing urban areas, the focus of urban renewal has likewise moved from large-scale physical reconstruction to fine-grained diagnosis and differentiated intervention [1]. Street space is the direct interface through which the public encounters the urban environment in everyday life, and visible characteristics such as greenery, openness, and walking conditions are associated with emotional experience [2]. However, objective environmental conditions and subjective emotional experiences are not always aligned. For example, in some low-density urban areas in the United Kingdom, newly built communities may provide favorable physical environments and public facilities, but residents’ subjective perceptions do not necessarily improve accordingly [3]. Conversely, some central urban areas may exhibit relatively low spatial quality but relatively positive public sentiment, which may be associated with well-developed supporting facilities and a high degree of functional mix [4]. Similarly, in some historic districts, relatively old physical environments may coexist with positive subjective experiences associated with social cohesion and emotional compensation [5]. These observations suggest that SSQ and public sentiment may form complex spatial patterns in particular locations. Identifying their spatial matching types can therefore help locate priority areas for improvement more accurately and provide a basis for street renewal and differentiated interventions.

Recent advances in urban big data and machine learning have created new opportunities to measure SSQ and public sentiment at the urban scale [6]. Street-view imagery continuously records the visible street environment from a pedestrian-oriented perspective, providing a data source for measuring street greenery, spatial form, pedestrian facilities, and street interfaces. Previous studies have shown that measurements derived from street-view imagery can reliably represent the physical street environment [7,8]. Semantic segmentation can further automate the processing of large image collections and identify and quantify multiple streetscape elements, thereby improving both the efficiency and spatial coverage of SSQ measurement [9,10].

Social media platforms contain opinions and emotional expressions generated during people’s everyday activities, providing abundant textual data for large-scale measurement of public sentiment [11]. Geographic information links textual content to specific locations, allowing dispersed emotional expressions to be aggregated within spatial units and revealing spatial differences in public sentiment [12,13]. However, the volume and complexity of social media text make consistent city-scale processing impractical through manual methods alone [14]. Natural language processing can classify positive and negative sentiment in batches and produce continuous sentiment scores, thereby improving processing efficiency and reproducibility [15,16].

Existing research suggests that SSQ is associated with public sentiment and that features such as street greenery, pedestrian environments, and spatial enclosure may correspond to different emotional responses [17,18]. At the same time, the degree of correspondence between the two varies across space, and some areas display a lack of synchronization between spatial quality and subjective experience [19,20]. Previous studies have compared objective environments with subjective experiences using matching classifications, indicator differences, or residuals and have identified marked spatial differentiation and clustering among different matching or divergence states [21,22,23]. Nevertheless, most studies have focused on the spatial distribution of matching types, objective–subjective discrepancies, or spatial relationships between relevant indicators. The correspondence between local matching states and neighborhood-level bivariate spatial associations remains comparatively underexplored.

To address this gap, this study integrates street-view imagery and Weibo data for the central urban area of Shanghai and develops a multilevel spatial diagnostic framework comprising same-grid matching identification, neighborhood-level bivariate spatial association, and cross-level comparison of types within 500 m regular grids. First, SSQ and public sentiment indicators are constructed to characterize their spatial differentiation and clustering. Second, their spatial matching relationships are identified at the same-grid and neighborhood levels, and the correspondence between the two diagnostic results is assessed. The stability of the matching types is then evaluated, and differences in their built environment characteristics are further analyzed on the basis of a comparison of five-dimensional SSQ profiles across stable types, thereby informing differentiated street renewal strategies. Accordingly, this study addresses three questions: (1) What spatial differentiation and clustering patterns characterize SSQ and public sentiment in the central urban area of Shanghai? (2) How do their matching patterns and spatial continuity differ between the same-grid and neighborhood levels? (3) Are the matching types stable, and what multidimensional quality characteristics and corresponding renewal strategies distinguish the stable types?

2. Materials and Methods

2.1. Area and Data Sources

2.1.1. Study Area and Spatial Analysis Units

Shanghai is a megacity on China’s eastern coast with a high concentration of population and urban functions. Its administrative area covers 6340.5 km², comprises 16 districts, and has approximately 24.8 million permanent residents [24]. In urban governance, the Shanghai Municipal Government has continued to promote the Beautiful Blocks initiative, strengthen human-centered services in urban space [25], and incorporate public satisfaction into urban performance evaluation [26]. Identifying the matching relationship between SSQ and public sentiment can therefore support a more comprehensive assessment of governance outcomes and differentiated renewal.

The Shanghai Master Plan 2017–2035 identifies the central urban area as the core concentration of diverse urban functions [27]. The area exhibits pronounced internal differences in land use, street form, and activity intensity. Accordingly, the central urban area of Shanghai was selected as the study area, providing a representative high-density urban setting for identifying SSQ–public sentiment matching types (Figure 1). Its dense population and public activities also generate continuous public expressions with temporal and spatial information, creating a basis for identifying spatial differences in sentiment. To harmonize the spatial scales of the street-view and Weibo datasets, all spatial data were projected to WGS 1984 and aggregated into 500 m × 500 m regular grids. The choice of this resolution was primarily based on the combined consideration of multi-source data integration and sample stability. Previous studies have used 500 m × 500 m grids to aggregate multi-source urban data, including street-view imagery, social media data, and built environment data [28,29,30]. It not only harmonizes the spatial scales of data from different sources but also closely corresponds to the local spatial range of residents’ daily walking activities, thereby helping to characterize the surrounding environmental features encountered during their daily activities [31,32,33]. Previous research has also shown that this resolution can reduce the data sparsity caused by excessively fine spatial divisions and provide a relatively stable basis for within-grid sample aggregation [34].

Figure 1. Study area. (a) Location of Shanghai in China; (b) location of the study area within Shanghai; (c) distribution of 500 m × 500 m analysis grids.

A total of 2642 candidate grids were generated. To improve sample stability, only grids containing at least 20 Weibo posts, at least five street-view samples, and complete key attributes were retained, resulting in 2085 valid analytical grids.

2.1.2. Geotagged Weibo Data

This study used geotagged Sina Weibo check-in data (https://weibo.com) to represent public sentiment. The public sentiment measure used in this study primarily captures the overall tendency of emotional expression among geotagged Weibo users across different locations. Given the selectivity associated with social media use and location-sharing behavior, the results are more likely to reflect the emotional expressions of relatively active users on the Weibo platform.

Weibo is one of China’s most widely used and active social media platforms, with approximately 600 million monthly active users and 200 million daily active users [35]. The text, timestamps, and spatial information associated with Weibo posts link public expressions to specific urban environments, and previous studies have demonstrated the suitability of Weibo data for public sentiment identification and spatiotemporal analysis [36,37]. The raw data included post text, publication time, geographic coordinates, and interaction information. Preprocessing involved removing HTML tags, URLs, and invalid characters; excluding invalid or excessively short texts; spatially filtering posts using their coordinates; and matching them to 500 m grids. After cleaning, 652,819 posts contained valid text and spatial information. To reduce instability in mean sentiment scores for grids with small samples, only grids with at least 20 posts were retained. The final analysis included 640,929 posts across 2085 valid grids.

2.1.3. Street-View Data

Street-view data were derived from Baidu Street View (https://api.map.baidu.com). Street-view images record visible elements such as roads, buildings, greenery, sky, and street facilities from a pedestrian-oriented perspective. They complement the limited ability of conventional two-dimensional spatial data to characterize street-level visual environments and support large-scale measurement of street space [38,39].

Using an OpenStreetMap (OSM) road network and road centerlines, a Python 3.7. program generated street-view sampling points at 50 m intervals. Corresponding images were then retrieved in batches through the Baidu Maps Open Platform API. Multiple horizontal views were obtained at each sampling point to reduce information loss associated with a single viewing direction. Image quality screening was subsequently conducted to remove sampling points without valid street-view coverage, and valid samples were assigned to 500 m grids using their coordinates. The 2085 valid analytical grids contained 174,630 street-view images, which were used to identify streetscape elements and calculate quality indicators.

2.1.4. Built Environment Data

To quantify the built environment characteristics of the study area, three types of spatial data were obtained: points of interest (POIs), road networks, and building footprints. POI and bus stop data were obtained from the AMap Open Platform (https://lbs.amap.com/). A Python program was used to retrieve POIs within the study area of Shanghai through the AMap Web Service API, including information on their names, categories, geographic coordinates, and addresses. Following data acquisition, duplicate records and records with evident positional offsets were removed or corrected. After data cleaning and spatial matching, the 2085 grids included in the analysis contained a total of 495,091 POI records and 2912 bus stops. The POIs were subsequently consolidated into functional categories based on the original AMap classification system.

Road network and building footprint data were both obtained from OpenStreetMap (OSM; https://www.openstreetmap.org/). The road network data primarily contained information on road locations, names, and classifications. These data were processed into topologically consistent road centerlines, from which road intersections were extracted. Within the 2085 grids included in the final analysis, the road centerlines had a total length of 6688.98 km, and 30,154 road intersections were identified. The building footprint data primarily contained building footprints and their spatial locations and were used in the subsequent analysis after being clipped to the study area and cleaned to remove duplicate features. After coordinate systems were harmonized, all the above datasets were matched to the 500 m × 500 m analytical grids.

2.2. Research Methods

2.2.1. Research Framework

This study first collected Baidu street-view imagery and geotagged Weibo data for the central urban area of Shanghai and matched the cleaned data to 500 m regular grids. SSQ and public sentiment were then quantified separately. A SegNet model was used to extract visual streetscape elements and construct quality indicators, while a Chinese RoBERTa model identified the positive sentiment probability of each Weibo post and generated grid-level sentiment scores. Univariate Global Moran’s I was subsequently used to test the spatial clustering of each variable, and Same-Grid Quadrant Classification identified Dual-Advantage, Sentiment-Lagging, Sentiment-Compensation, and Dual-Low zones. At the neighborhood level, Global Bivariate Moran’s I and Bivariate LISA were used to assess the spatial association between SSQ in a focal grid and public sentiment in neighboring grids, after which the LISA results were cross-tabulated with the same-grid types. Mean-based threshold replacement and Bootstrap resampling were then used to identify stable matching types. Finally, five-dimensional SSQ profiles and built environment characteristics were compared across the stable types to inform differentiated street renewal recommendations (Figure 2).

Figure 2. Overview of the research workflow.

2.2.2. Construction of the SSQ Index

Drawing on the urban design quality framework proposed by Ewing and Handy, including enclosure, human scale, transparency, and complexity [40], this study links SSQ to observable physical elements and operationalizes these concepts using visual elements identifiable through street-view semantic segmentation. Five dimensions were constructed: visual naturalness, spatial openness, pedestrian friendliness, pedestrian–vehicle interaction, and interface permeability (Table 1). These dimensions were represented by nine quantitative indicators: green view index, sky visibility, building visibility, sidewalk visibility, road visibility, pedestrian presence, car visibility, wall visibility, and fence visibility.

The indicators were extracted using the SegNet Semantic Segmentation Model, which has an encoder–decoder architecture [41]. The encoder extracts semantic image features, while the decoder restores spatial information to achieve pixel-level classification of street-view images. Weights pre-trained on the ADE20K Dataset were used [42]. The numbers of pixels corresponding to vegetation, sky, buildings, sidewalks, roads, pedestrians, cars, walls, and fences were calculated and divided by the total number of pixels in each street-view image to produce the nine streetscape visual indicators.

Table 1. SSQ Indicator System.

Dimension	Indicator and Direction	Meaning	Calculation Formula	Reference
Visual Naturalness	Green View Index (+)	The visible proportion of green vegetation in street-view images, reflecting the coverage of natural elements.	${\mathrm{GVI}}_n={\displaystyle \frac{G_n}{A_n}}$	[43,44]
Spatial Openness	Sky Visibility (+)	The visibility of the sky in street-view images, reflecting the visual openness of street space.	${\mathrm{SV}}_n={\displaystyle \frac{S_n}{A_n}}$	[10,45]
	Building Visibility (−)	The visual occupation of street space by building façades, reflecting spatial enclosure pressure.	${\mathrm{BV}}_n={\displaystyle \frac{B_n}{A_n}}$	[46,47]
Pedestrian Friendliness	Sidewalk Visibility (+)	The visibility of sidewalks and other pedestrian spaces in street-view images.	${\mathrm{SW}}_n={\displaystyle \frac{W_n}{A_n}}$	[48,49]
	Road Visibility (−)	The visual occupation of street space by motor-vehicle roadways.	${\mathrm{RV}}_n={\displaystyle \frac{R_n}{A_n}}$	[50,51]
Pedestrian–Vehicle Interaction	Pedestrian Presence (+)	The presence of pedestrians in street-view images, reflecting the level of street activity.	${\mathrm{PV}}_n={\displaystyle \frac{P_n}{A_n}}$	[52,53]
	Car Visibility (−)	The visibility of cars in street-view images, reflecting motor-vehicle interference.	${\mathrm{CV}}_n={\displaystyle \frac{C_n}{A_n}}$	[54,55]
Interface Permeability	Wall Visibility (−)	The visibility of solid walls in street-view images, reflecting interface enclosure and visual obstruction.	${\mathrm{WV}}_n={\displaystyle \frac{M_n}{A_n}}$	[56,57]
	Fence Visibility (−)	The visibility of fences in street-view images, reflecting spatial separation and constraints on interface permeability.	${\mathrm{FV}}_n={\displaystyle \frac{F_n}{A_n}}$	[49,58]

Note: $A_n$ denotes the total number of pixels in street-view image $n$. $G_n$, $S_n$, $B_n$, $W_n$, $R_n$, $P_n$, $C_n$, $M_n$, and $F_n$ denote the numbers of pixels classified as green vegetation, sky, buildings, sidewalks, roads, pedestrians, cars, walls, and fences, respectively. After the indicators were calculated for each image, the values of all street-view samples within the same grid were averaged. In addition, “+” and “−” indicate the direction assigned to each indicator within the evaluation framework used in this study; the specific effects of these indicators may vary across spatial environments and contexts of use.

Table 1. SSQ Indicator System.

Dimension	Indicator and Direction	Meaning	Calculation Formula	Reference
Visual Naturalness	Green View Index (+)	The visible proportion of green vegetation in street-view images, reflecting the coverage of natural elements.	${\mathrm{GVI}}_n={\displaystyle \frac{G_n}{A_n}}$	[43,44]
Spatial Openness	Sky Visibility (+)	The visibility of the sky in street-view images, reflecting the visual openness of street space.	${\mathrm{SV}}_n={\displaystyle \frac{S_n}{A_n}}$	[10,45]
	Building Visibility (−)	The visual occupation of street space by building façades, reflecting spatial enclosure pressure.	${\mathrm{BV}}_n={\displaystyle \frac{B_n}{A_n}}$	[46,47]
Pedestrian Friendliness	Sidewalk Visibility (+)	The visibility of sidewalks and other pedestrian spaces in street-view images.	${\mathrm{SW}}_n={\displaystyle \frac{W_n}{A_n}}$	[48,49]
	Road Visibility (−)	The visual occupation of street space by motor-vehicle roadways.	${\mathrm{RV}}_n={\displaystyle \frac{R_n}{A_n}}$	[50,51]
Pedestrian–Vehicle Interaction	Pedestrian Presence (+)	The presence of pedestrians in street-view images, reflecting the level of street activity.	${\mathrm{PV}}_n={\displaystyle \frac{P_n}{A_n}}$	[52,53]
	Car Visibility (−)	The visibility of cars in street-view images, reflecting motor-vehicle interference.	${\mathrm{CV}}_n={\displaystyle \frac{C_n}{A_n}}$	[54,55]
Interface Permeability	Wall Visibility (−)	The visibility of solid walls in street-view images, reflecting interface enclosure and visual obstruction.	${\mathrm{WV}}_n={\displaystyle \frac{M_n}{A_n}}$	[56,57]
	Fence Visibility (−)	The visibility of fences in street-view images, reflecting spatial separation and constraints on interface permeability.	${\mathrm{FV}}_n={\displaystyle \frac{F_n}{A_n}}$	[49,58]

Note: $A_n$ denotes the total number of pixels in street-view image $n$. $G_n$, $S_n$, $B_n$, $W_n$, $R_n$, $P_n$, $C_n$, $M_n$, and $F_n$ denote the numbers of pixels classified as green vegetation, sky, buildings, sidewalks, roads, pedestrians, cars, walls, and fences, respectively. After the indicators were calculated for each image, the values of all street-view samples within the same grid were averaged. In addition, “+” and “−” indicate the direction assigned to each indicator within the evaluation framework used in this study; the specific effects of these indicators may vary across spatial environments and contexts of use.

Because the nine streetscape indicators differ in their value ranges and directional relationships with SSQ, Min–Max normalization was applied across the 2085 valid grids to standardize the indicators and align their directions, as follows:

$$z_{ik}={\displaystyle \frac{s_k\left(x_{ik}-x_k^{\min }\right)+\left(1-s_k\right)\left(x_k^{\max }-x_{ik}\right)}{x_k^{\max }-x_k^{\min }}}$$

(1)

where $z_{ik}$ denotes the standardized value of the $k$-th indicator in grid $i$; $x_{ik}$ denotes the corresponding original value; $x_k^{\min }$ and $x_k^{\max }$ denote the minimum and maximum values, respectively, of the $k$-th indicator across all valid grids; and $s_k$ is the indicator direction parameter, which is set to 1 for positive indicators and 0 for negative indicators. After processing, all indicator values range from 0 to 1, with higher values indicating higher SSQ, as reflected by the corresponding indicator.

On this basis, the score for each dimension was calculated as the arithmetic mean of its constituent indicators, while the overall SSQ score was calculated as the arithmetic mean of the nine standardized indicators, as follows:

$$SS{Q}_i={\displaystyle \frac{1}{9}}{\sum }_{k=1}^9{z}_{ik}$$

(2)

where $SS{Q}_i$ denotes the overall SSQ score for grid $i$, and $z_{ik}$ denotes the standardized value of the $k$-th indicator in grid $i$. $SS{Q}_i$ ranges from 0 to 1, with higher values indicating higher overall SSQ in the corresponding grid.

2.2.3. Measurement of Public Sentiment

A RoBERTa model was used to classify sentiment in Weibo text. Specifically, this study applied the uer/roberta-base-finetuned-jd-binary-chinese binary classification model fine-tuned on Chinese-language data. Built on Chinese RoBERTa, the model was pre-trained on the CLUECorpusSmall corpus and fine-tuned for positive–negative sentiment classification using the JD binary Chinese review dataset [59]. RoBERTa retains BERT’s bidirectional Transformer encoder architecture and generates contextualized semantic representations by jointly considering the left and right contexts of each token [60]. Compared with lexicon-based methods that rely on fixed sentiment values and predefined rules, contextualized models are better suited to handling context-dependent Chinese word meanings and irregular expressions in online text [61].

The public sentiment indicator was constructed in two stages: sentiment classification of individual Weibo posts and within-grid aggregation. First, each cleaned post was entered into the model, which produced negative and positive logits. The Softmax function was then used to calculate the positive sentiment probability for Weibo post $i$, as follows:

$$P_i^{+}={\displaystyle \frac{exp\left(z_i^{+}\right)}{exp\left(z_i^{-}\right)+exp\left(z_i^{+}\right)}}$$

(3)

where $z_i^{-}$ and $z_i^{+}$ are the negative and positive logits, respectively, for Weibo post $i$. This study uses $P_i^{+}$ as positive sentiment score for an individual post. The score ranges from 0 to 1, with higher values indicating more positive sentiment.

The positive sentiment probabilities of all valid posts within each 500 m grid were then averaged arithmetically to obtain the grid-level public sentiment score:

$$P{S}_g={\displaystyle \frac{1}{n_g}}{\displaystyle \sum _{i=1}^{n_g}}{P}_i^{+}$$

(4)

where $P{S}_g$ is the public sentiment score for grid $g$, and $n_g$ is the number of valid Weibo posts in the grid. To reduce instability in the mean sentiment scores of small-sample grids and to ensure adequate street-view coverage and complete attributes, only grids containing at least 20 valid posts and meeting the remaining data-completeness criteria were retained for subsequent analysis.

To examine the applicability of the model to the Weibo texts used in this study, a manual consistency check was conducted on the sentiment classification results. Based on the positive sentiment probabilities generated by the model, a total of 210 Weibo posts were sampled from the high-score, low-score, and near-threshold ranges around 0.5, and their sentiment polarity was manually assessed according to the semantic meaning of the text. Of these, 174 posts with clearly identifiable positive or negative sentiment polarity were used to assess agreement with the model classifications. The observed agreement rate between the manual judgments and model classifications was 82.76%, suggesting that the model had reasonable applicability for identifying positive and negative sentiment in the Weibo texts analyzed in this study.

2.2.4. Spatial Autocorrelation and Bivariate Spatial Association Analysis

Spatial autocorrelation analysis evaluates spatial dependence in geographic variables by testing whether their spatial distributions deviate from randomness and whether they exhibit overall clustering or dispersion [62]. Bivariate spatial association analysis measures the direction and strength of the relationship between the value of one variable in a focal spatial unit and another variable in neighboring units [63]. This study applied global and local spatial statistics to examine the relationship between SSQ and public sentiment from the perspectives of univariate spatial clustering and bivariate neighborhood association.

Because each of the spatial statistics described above requires a definition of spatial neighborhood, a common spatial weights matrix was first constructed. Specifically, first-order Queen contiguity was adopted, whereby two grids were defined as neighbors if they shared either an edge or a vertex. The initial contiguity weight was assigned a value of 1 for neighboring grids and 0 otherwise. This definition incorporates directly adjacent grids in the horizontal, vertical, and diagonal directions, thereby providing a relatively complete representation of the local neighborhood surrounding each 500 m × 500 m regular grid. The spatial weights matrix was then row-standardized so that the sum of the contiguity weights for each non-isolated grid equaled 1. Univariate Global Moran’s I, Global Bivariate Moran’s I, and Bivariate LISA were all calculated using this spatial weights matrix, with statistical significance consistently assessed at p < 0.05.

First, Univariate Global Moran’s I was calculated separately for SSQ and public sentiment to evaluate their overall spatial autocorrelation:

$$I={\displaystyle \frac{N}{S_0}}\cdot {\displaystyle \frac{{\displaystyle {\sum }_{i=1}^N}{\displaystyle {\sum }_{j=1}^N}{w}_{ij}(x_i-\overline{x})(x_j-\overline{x})}{{\displaystyle {\sum }_{i=1}^N}(x_i-\overline{x}{)}^2}}$$

(5)

where $N$ is the number of grids; $x_i$ and $x_j$ are the observed values in grids $i$ and $j$, respectively; $\overline{x}$ is the mean of the variable; $w_{ij}$ is the row-standardized first-order Queen contiguity weight; and $S_0={\sum }_{i=1}^N{\sum }_{j=1}^N{w}_{ij}$ is the sum of all spatial weights. A positive Moran’s I indicates that grids with similar values tend to cluster, whereas a negative value indicates that high- and low-value grids tend to be adjacent. A value close to the expectation under spatial randomness indicates weak overall spatial autocorrelation.

Building on this analysis, Global Bivariate Moran’s I was used to test the overall spatial association between SSQ in a focal grid and public sentiment in neighboring grids:

$$I_{QP}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^N}{z}_{Q_i}\left({\displaystyle {\sum }_{j=1}^N}{w}_{ij}|z_{P_j}\right)}{{\displaystyle {\sum }_{i=1}^N}{z}_{Q_i}^2}}$$

(6)

where $z_{Q_i}$ is the standardized SSQ value of focal grid $i$; $z_{P_j}$ is the standardized public sentiment value of neighboring grid $j$; ${\displaystyle {\sum }_{j=1}^N}{w}_{ij}{z}_{P_j}$ is the spatial lag of public sentiment around grid $i$. $I_{QP}>0$ indicates an overall same-direction spatial association between SSQ and neighboring public sentiment, whereas $I_{QP}<0$ indicates an overall reverse spatial association.

Bivariate LISA was then used to identify local spatial association types:

$$I_i^{QP}=z_{Q,i}{\displaystyle \sum _{j=1}^N}{w}_{ij}{z}_{P,j}$$

(7)

Based on combinations of SSQ in the focal grid and the spatially weighted mean of public sentiment in neighboring grids, statistically significant grids were classified into four association types, while grids that did not pass the significance test were classified as Not Significant (NS). Global Bivariate Moran’s I and Bivariate LISA were assessed using 999 random permutations, with p-value below 0.05 indicating statistical significance.

2.2.5. Spatial Matching and Robustness Analysis

To identify combinations of SSQ and public sentiment at different spatial levels, this study examined same-grid matching and neighborhood spatial association and evaluated the robustness of the same-grid matching types through threshold replacement and Bootstrap resampling.

At the same-grid level, Same-Grid Quadrant Classification was conducted with SSQ on the horizontal axis and public sentiment on the vertical axis. The medians of the two variables across the 2085 valid grids were used as the high–low thresholds. The grids were classified into four types: Dual-Advantage Zone (HH), Sentiment-Lagging Zone (HL), Sentiment-Compensation Zone (LH), and Dual-Low Zone (LL). The first letter indicates the relative level of SSQ in the focal grid, and the second indicates the relative level of public sentiment in that same grid. This classification identifies same-grid matching and mismatch between SSQ and public sentiment.

At the neighborhood level, the Bivariate Local Indicators of Spatial Association (Bivariate LISA) results were used to identify local spatial associations between SSQ in each focal grid and public sentiment in neighboring grids.

To assess the sensitivity of Same-Grid Quadrant Classification to threshold selection, the mean SSQ and public sentiment scores were substituted for their medians, and the HH, HL, LH, and LL types were recalculated and compared with the main results. A grid whose type remained unchanged after threshold replacement was considered relatively insensitive to the high–low classification criterion.

Within-grid Bootstrap resampling was then used to assess the sensitivity of public sentiment measurement and same-grid classification to variation in the Weibo sample. For each grid, 500 samples were drawn with replacement while retaining the original number of posts; the mean public sentiment score and the associated high–low threshold were recalculated in each iteration, and the quadrant type was reassigned. A grid was classified as stable when its type remained unchanged under mean-based threshold replacement and its original type was retained in at least 80% of the Bootstrap resamples. Stable grids were further classified as Stable Dual-Advantage Zones (ST-HH), Stable Sentiment-Lagging Zones (ST-HL), Stable Sentiment-Compensation Zones (ST-LH), or Stable Dual-Low Zones (ST-LL); all remaining grids were classified as Uncertain Zones (UNCERT).

2.2.6. Comparison of Built Environment Characteristics Across Stable Spatial Types

To further explore differences in the built environment characteristics associated with different stable spatial types, this study drew on multidimensional measurement frameworks used in previous built environment and urban studies [64,65,66] and considered the characteristics of Shanghai’s high-density urban environment. The built environment variables were grouped into five dimensions: functional configuration, development intensity, road network structure, public transport, and daily service and activity facilities. All built environment variables were calculated using 500 m × 500 m grids to maintain consistency with the analytical units used for SSQ and public sentiment. The built environment variables and their calculation methods are presented in Table 2.

Table 2. Built Environment Variables and Calculation Methods.

Dimension	Variable	Calculation formula	Reference
Functional Configuration	Land-Use Mix	$M_i=-{\sum }_{k=1}^K{p}_{ik}\ln \left(p_{ik}\right)$	[67,68]
Development Intensity	Building Coverage Ratio	$BCR=\Sigma Abi/Sgrid$	[69,70]
Road Network Structure	Road Network Density	$RN{D}_i=L_R,i/A_i$	[71,72]
	Intersection Density	$I{D}_i=N_I,i/A_i$	[73,74]
Public Transport	Bus Stop Density	$BS{D}_i=N_{B}S,i/A_i$	[75,76]
Daily Service and Activity Facilities	Retail Facility Density	$RF{D}_i=N_{R}F,i/A_i$	[65,77]
	Daily Service Facility Density	$DSF{D}_i=N_{D}SF,i/A_i$	[19,78]
	Leisure Facility Density	$LF{D}_i=N_{L}F,i/A_i$	[79,80]

This study first applied the Kruskal–Wallis H test to compare the overall differences in each built environment variable among the four stable types: ST-HH, ST-HL, ST-LH, and ST-LL. On this basis, two prespecified pairwise comparisons were conducted using two-sided Mann–Whitney U tests: ST-HL versus ST-HH and ST-LH versus ST-LL. The two comparisons were conducted separately within the relatively high- and low-SSQ categories, allowing differences in built environment characteristics between each stable mismatch type and its corresponding matching type to be examined at the same relative SSQ level. Because multiple built environment variables were examined simultaneously, the Benjamini–Hochberg method was applied separately to adjust the p-values from the overall Kruskal–Wallis tests and from each of the two prespecified comparisons. Statistical significance was determined at q < 0.05. Continuous variables in each group were summarized using the median and interquartile range (IQR).

3. Results

3.1. Spatial Distribution and Clustering Characteristics of SSQ and Public Sentiment

3.1.1. Overall Patterns of SSQ and Public Sentiment

Overall, SSQ was relatively concentrated across grids. Public sentiment likewise showed limited overall variation, although discernible differences remained between grids. Among the five SSQ dimensions, Interface Permeability had the highest mean (0.8682), whereas Pedestrian Friendliness had the lowest (0.3094). This suggests that street interfaces were generally permeable, while pedestrian space remained comparatively limited. Visual Naturalness had the largest SD (0.1494), followed by Spatial Openness (0.1184), indicating greater inter-grid variation in visible greenery and spatial openness than in the other dimensions (Table 3).

Table 3. Descriptive Statistics of the Main Variables.

Variable	N	Minimum	Maximum	Mean	Median	SD
Visual Naturalness	2085	0.0000	1.0000	0.3203	0.2994	0.1494
Spatial Openness	2085	0.2438	0.9880	0.7039	0.7226	0.1184
Pedestrian Friendliness	2085	0.0686	0.8986	0.3094	0.2927	0.0991
Pedestrian–Vehicle Interaction	2085	0.0122	0.9254	0.3527	0.3604	0.0842
Interface Permeability	2085	0.1833	1.0000	0.8682	0.8847	0.0776
Street Space Quality	2085	0.4087	0.6961	0.5321	0.5326	0.0347
Public Sentiment	2085	0.5279	0.9306	0.7480	0.7495	0.0397

3.1.2. Spatial Differentiation of SSQ

Overall, SSQ displayed pronounced patchiness and spatial heterogeneity (Figure 3). Medium-value grids predominated, while high- and low-value patches were interspersed and formed several locally contiguous areas, producing a multicentric clustering pattern. The five dimensions showed distinct spatial distributions. Visual Naturalness was dominated by scattered small patches with interwoven high and low values. High Pedestrian Friendliness values were more concentrated and formed clearer contiguous areas. Spatial Openness showed the clearest differentiation, with low values clustered in the center and relatively high values surrounding them. Interface Permeability and Pedestrian–Vehicle Interaction exhibited multi-area clustering, with high- and low-value patches interspersed throughout the study area.

Figure 3. Spatial Distribution of Multidimensional SSQ.

3.1.3. Spatial Differentiation of Public Sentiment and Weibo Sample Counts

Public sentiment displayed a comparatively smooth spatial pattern dominated by medium-value grids, with scattered local clusters of high and low values (Figure 4). Several small high-value clusters occurred mainly in the central and southern parts of the study area, whereas low-value grids were interspersed across different areas; large contiguous high- or low-value areas were limited. By contrast, Weibo sample counts showed a clearer spatial gradient. Grids with high sample counts were concentrated from the central to southwestern portions of the central urban area and gradually transitioned toward medium- and low-count grids at the periphery. Counts were relatively low in northern, eastern, and edge areas. Thus, Weibo activity was spatially more intensive in the central and southwestern portions of Shanghai’s central urban area and declined outward from the main cluster.

Figure 4. Spatial Distribution of Public Sentiment and Weibo Sample Counts.

3.1.4. Global Spatial Autocorrelation of SSQ and Public Sentiment

Building on the spatial distribution analysis, Univariate Global Moran’s I was used to test the spatial clustering of SSQ and public sentiment separately (Table 4). Both variables exhibited significant positive spatial autocorrelation, with Moran’s I values of 0.3285 and 0.1444, respectively (p < 0.001), indicating that grids with similar values were significantly more likely to be spatially adjacent. Positive spatial autocorrelation was stronger for SSQ, whereas public sentiment exhibited comparatively weaker spatial clustering.

Table 4. Global Spatial Autocorrelation Results for SSQ and public sentiment.

Variable	Moran’s I	z-Score	p-Value	Spatial Pattern
Street Space Quality	0.3285	26.968	<0.001	Significant positive spatial autocorrelation
Public Sentiment	0.1444	11.875	<0.001	Significant positive spatial autocorrelation

3.2. Spatial Matching Patterns of SSQ and Public Sentiment

3.2.1. Spatial Distribution of Matching Types

To identify same-grid SSQ–public sentiment matching types, the medians of the two variables were used as thresholds to classify grids into HH, HL, LH, and LL combinations (Figure 5). Each quadrant contained a substantial number of grids, indicating that SSQ and public sentiment occurred in multiple combinations rather than being dominated by a single type. The two mismatch types, HL and LH, together accounted for 52.90% of all grids.

Figure 5. Same-Grid Quadrant Scatterplot of SSQ and Public Sentiment. The red dashed vertical and horizontal lines indicate the median street space quality score (0.5326) and the median public sentiment score (0.7495), respectively, which were used as the thresholds for classifying the grids into the HH, HL, LH, and LL types.

The combination types exhibited clear regional heterogeneity (Figure 6). Areas within the Inner Ring Road and in the western portion of the Middle Ring Road were characterized mainly by interspersed LL and LH grids. SSQ was generally low in these areas, whereas public sentiment varied, with relatively positive sentiment occurring in scattered locations. HL grids were more concentrated in the eastern portion of the Middle Ring Road and southeast near the Outer Ring Road, where they formed local linear or contiguous patches, indicating a more pronounced mismatch between high SSQ and low public sentiment. HL grids were relatively prevalent near the Outer Ring Road but were more spatially dispersed. HH grids were generally scattered and formed only a few small contiguous patches, suggesting that the Dual-Advantage pattern covered a limited area and did not form extensive continuous zones.

Figure 6. Spatial Distribution of Same-Grid Matching Types.

3.2.2. Neighborhood Clustering Characteristics of Spatial Matching Patterns

Building on the identification of same-grid matching types, this study further examined whether the mismatch relationship extended to the neighborhood level. Global Bivariate Moran’s I was −0.0417 (p = 0.002), indicating a significant negative spatial association between SSQ in focal grids and public sentiment in neighboring grids (Figure 7). Although the association was weak, it was statistically supported by the random permutation test, indicating a non-random but limited reverse spatial relationship at the study-area level.

Figure 7. Global Bivariate Moran’s I Scatterplot of SSQ and Neighboring Public Sentiment. The red solid line represents the fitted regression line, whose slope corresponds to the Global Bivariate Moran’s I. The black dashed vertical and horizontal lines indicate zero values of standardized SSQ and the spatial lag of standardized public sentiment, respectively, dividing the scatterplot into the HH, HL, LH, and LL quadrants.

At the local level, SSQ and neighboring public sentiment showed a mixture of matching and mismatch patterns rather than universal correspondence (Figure 8). Among statistically significant grids, the mismatch types LISA-HL and LISA-LH accounted for 58.25%, exceeding the combined share of LISA-HH and LISA-LL. Local association types varied among ring zones. Within the Inner Ring Road, LISA-HH and LISA-LH were concentrated in central–western and southern areas and formed several interwoven clusters. Significant types in the Middle Ring zone were more scattered and mainly represented local extensions of LISA-HH and LISA-LH. Around the Outer Ring Road, directional differentiation was more pronounced: LISA-HH clusters occurred mainly in the north, while LISA-HL dominated the east and southeast in linear or patch-like formations, with some areas adjoining LISA-LL. Significant grids were comparatively sparse along the western and southern margins.

Figure 8. Bivariate LISA Cluster Map of SSQ and Neighboring Public Sentiment.

3.2.3. Locality and Continuity of Spatial Matching Patterns

To determine whether same-grid matching relationships were accompanied by significant neighborhood associations in the same direction, the Same-Grid Quadrant types were cross-tabulated with the Bivariate LISA results (Figure 9). Only 19.18% of all grids exhibited significant Bivariate LISA associations. Thus, most same-grid matches and mismatches remained local and did not form statistically significant neighborhood-level spatial associations.

Figure 9. Composition and Significance Rates of Bivariate LISA Types across Same-Grid SSQ–Public Sentiment Combinations. (a) composition of bivariate LISA cluster types among significant grids within each same-grid combination; (b) proportion of grids with significant bivariate LISA results within each same-grid combination. The gray dashed vertical line in (b) indicates the overall proportion of significant bivariate LISA grids.

Among same-grid mismatches that did exhibit significant neighborhood associations, directional agreement was high. Of the HL and LH grids assigned to a significant LISA category, 78.8% and 84.0%, respectively, corresponded to LISA-HL and LISA-LH. When a same-grid mismatch was accompanied by a significant neighborhood association, the sentiment direction of the focal grid generally extended to its neighbors: HL grids tended to be surrounded by grids with low public sentiment, whereas LH grids tended to be surrounded by grids with high public sentiment. Matching types showed a similar pattern. The proportion of HH grids associated with same-direction LISA-HH (18.53%) was higher than the proportion of LL grids associated with LISA-LL (7.54%).

3.3. Stable-Type Identification and Multidimensional Quality Profiles

3.3.1. Spatial Distribution of Stable Matching Types

To evaluate the stability of the quadrant classifications, mean-based threshold replacement and 500 within-grid Bootstrap resamples were conducted. The classifications were generally robust to threshold changes, although some grids remained sensitive to sample variation. Together, the two tests identified 1307 stable grids (62.69% of all grids). Stable mismatch types accounted for 33.38%, indicating that some mismatch grids retained a certain degree of stability under changes in the classification threshold and sample variation.

Grids were categorized as Stable High–High (ST-HH), Stable Low–Low (ST-LL), stable mismatch types (ST-HL and ST-LH), or uncertain areas (UNCERT). The types formed interwoven spatial patches (Figure 10). Within the Inner Ring Road, ST-LH and ST-LL were interspersed; ST-LH was relatively concentrated in central-western and south-central areas, ST-LL occurred mainly across northern and eastern areas, and ST-HH formed only a few local patches in the south. The Middle Ring zone had a mixed composition: ST-LH and ST-LL remained more common in the west and southwest, while ST-HL increased toward the east. Around the Outer Ring Road, an east–west contrast emerged. ST-HL was concentrated in the east and southeast, where it formed several continuous or semi-continuous patches, whereas ST-LL occurred more frequently along the northern, western, and southern margins. ST-HH appeared in all ring zones but remained scattered. UNCERT grids were interspersed widely among the stable types, indicating substantial classification uncertainty in some locations.

Figure 10. Spatial Distribution of Stable SSQ–Public Sentiment Combination Types.

3.3.2. Multidimensional SSQ Differences Among Stable Combination Types

Dimension-wise standardized z-scores were used to compare the five-dimensional SSQ profiles of the stable types (Figure 11). The two stable mismatch types exhibited different multidimensional structures. ST-HL had its strongest advantage in Spatial Openness, followed by Pedestrian–Vehicle Interaction, whereas Pedestrian Friendliness was comparatively weakest. ST-LH showed a slightly above-average level only for Pedestrian Friendliness; all other dimensions were below average, with Spatial Openness and Pedestrian–Vehicle Interaction lowest, followed by Visual Naturalness. By comparison, ST-HH exhibited clear advantages across all five dimensions, whereas ST-LL showed multidimensional deficits. The two stable mismatch types therefore require differentiated diagnosis and optimization strategies.

Figure 11. Five-Dimensional SSQ Profiles across Stable Matching Types.

3.4. Differences in Built Environment Characteristics Across Stable Matching Types

3.4.1. Overall Differences Among the Four Stable Matching Types

Between-group comparisons of built environment characteristics were conducted using the 1307 grids whose classifications remained stable in the Bootstrap robustness test. The Kruskal–Wallis H test indicated significant overall differences among ST-HH, ST-HL, ST-LH, and ST-LL for all eight built environment variables. On this basis, prespecified pairwise comparisons of ST-HL versus ST-HH and ST-LH versus ST-LL were conducted to further identify differences in built environment characteristics within these two pairs of stable matching types (Table 5).

Table 5. Adjusted q-Values for Nonparametric Comparisons of Built Environment Characteristics across Stable Matching Types.

Built Environment Variable	Kruskal–Wallis (q)	ST-HL vs. ST-HH (q)	ST-LH vs. ST-LL (q)
Land-Use Mix	<0.001	0.013	<0.001
Building Coverage Ratio	<0.001	0.210	<0.001
Road Network Density	<0.001	0.241	0.665
Intersection Density	<0.001	0.058	0.011
Bus Stop Density	0.043	0.213	0.301
Retail Facility Density	<0.001	0.002	<0.001
Daily Service Facility Density	<0.001	0.013	<0.001
Leisure Facility Density	<0.001	0.002	<0.001

Note: q-values are Benjamini–Hochberg-adjusted p-values. Statistical significance was defined as q < 0.05.

3.4.2. Between-Group Differences at Different SSQ Levels

Overall, the differences common to both prespecified comparisons were concentrated in land-use mix and the provision of daily service and activity facilities. ST-HL exhibited lower values than ST-HH for all these indicators, whereas ST-LH exhibited higher values than ST-LL, indicating a relatively consistent pattern of differentiation in these built environment characteristics across the stable matching types. By contrast, significant differences in building coverage ratio and intersection density occurred only between ST-LH and ST-LL (Figure 12).

Figure 12. Distributions of Built Environment Characteristics across Stable Matching Types.

Specifically, among the two types of grids characterized by relatively high SSQ, the differences between ST-HL and ST-HH were mainly observed in land-use mix and the provision of daily service and activity facilities. Land-use mix, retail facility density, daily service facility density, and leisure facility density were all significantly lower in ST-HL than in ST-HH q < 0.05, whereas building coverage ratio, road network density, intersection density, and bus stop density showed no significant differences.

Among the two types of grids characterized by relatively low SSQ, the differences between ST-LH and ST-LL encompassed a broader range of built environment characteristics. Land-use mix, building coverage ratio, intersection density, retail facility density, daily service facility density, and leisure facility density were all significantly higher in ST-LH than in ST-LL q < 0.05, whereas road network density and bus stop density showed no significant differences.

4. Discussion

4.1. Spatial Mismatch Between SSQ and Public Sentiment Across Analytical Levels

In Shanghai’s high-density urban environment, SSQ and public sentiment did not exhibit a consistently synchronized relationship, but instead exhibited different mismatch states. Although each variable showed significant spatial clustering, relatively high SSQ did not necessarily correspond to more positive public sentiment, and vice versa. This finding is consistent with previous research on relationships between environmental conditions and sentiment. Li et al. reported that higher spatial integration was not accompanied by a correspondingly stronger perception of historic character [81]. Gong et al. identified a spatial mismatch between urban park accessibility and residents’ subjective well-being [82]. Similarly, Kang et al. found spatial divergence between ecological conditions and visual perception [83].

The two mismatch types exhibited different spatial distributions within the central urban area. Grids characterized by relatively high SSQ but relatively low public sentiment were concentrated mainly in the eastern and southeastern parts of central Shanghai. By contrast, grids characterized by relatively low SSQ but relatively high public sentiment were concentrated primarily in the central-western and south-central areas. Further comparison of built environment characteristics showed that, among the two types of grids with relatively high SSQ, those with higher public sentiment exhibited higher levels of land-use mix and provision of daily service and activity facilities. Previous studies suggest that higher land-use mix and a richer provision of daily service facilities are generally associated with a greater diversity of activity destinations and may provide residents with more opportunities for daily travel, leisure activities, and social interaction [84,85,86]. Among the two types of grids with relatively low SSQ, differences in built environment characteristics were more varied. In addition to land-use mix and the provision of daily service and activity facilities, significant differences were also observed in development intensity and road network structure. Previous studies suggest that greater road network connectivity and more diverse route choices may provide residents with a more convenient environment for daily travel [87,88,89].

Furthermore, by integrating street-view and Weibo big data with quadrant classification, Bivariate LISA, and cross-level comparison, this study developed a multilevel spatial diagnostic framework that extends from identifying same-grid combinations to testing neighborhood clustering. Most same-grid mismatches were not accompanied by significant same-direction neighborhood associations, indicating the localized nature of these mismatch states. However, among the small subset of mismatched grids that formed significant neighborhood clusters, the original mismatch direction generally persisted, indicating a degree of neighborhood continuity in the direction of public sentiment. In addition, some mismatched grids retained their original types after threshold replacement and Bootstrap resampling, indicating a degree of classification stability.

4.2. Differentiated Street Renewal Implications Based on Stable Types

Urban spatial governance in Shanghai is shifting from incremental expansion toward quality improvement of existing urban spaces. In the context of high-density development, intensive functional mixing, and limited space for new construction, street governance should move beyond the continued addition of environmental elements and identify specific SSQ–public sentiment types to establish a renewal logic of type-based diagnosis and differentiated intervention. Accordingly, this study proposes differentiated renewal strategies in terms of mismatch correction, tiered quality enhancement, and multi-scale coordination.

Mismatch correction: The two mismatch types exhibited different built environment characteristics and should therefore be addressed through differentiated renewal approaches based on their existing conditions. For areas with high SSQ and low public sentiment, while maintaining the existing level of SSQ, attention may be extended to land-use mix and the provision of daily service and activity facilities, with the configuration of relevant functions and service facilities optimized according to actual needs. For areas with low SSQ and high public sentiment, improvements in SSQ should be accompanied by attention to and retention of the existing land-use mix, daily service provision, and road network conditions. Low-cost, incremental interventions can then be used to address deficiencies in greenery, openness, and other aspects of street space, thereby aligning environmental improvements with established everyday needs.

Tiered quality enhancement: The two matching types imply different resource-allocation strategies. Areas with high SSQ and high public sentiment should emphasize maintenance and enhancement, consolidating existing SSQ and service conditions through routine maintenance and fine-grained management while avoiding redundant construction. Areas with low SSQ and low public sentiment require comprehensive improvement. Alongside improvements in SSQ, the provision of relevant facilities should be enhanced according to the functional configuration and daily service needs of each area, thereby promoting coordinated improvements in the street environment and functional service conditions.

Multi-scale coordination: Because most mismatches are localized, node- or street-segment-level micro-renewal is generally appropriate. Areas that form contiguous clusters require coordination at the street-corridor or area level. In places with unstable classifications, field validation and public feedback should be incorporated into a long-term monitoring mechanism, with SSQ and public sentiment assessments updated periodically. Intervention directions should be determined only after the type and its trajectory become sufficiently stable.

4.3. Limitations and Future Research Directions

Although this study integrates street-view and Weibo big data within a multilevel framework, several limitations remain. First, social media data provide an effective means of spatially quantifying public sentiment, but their user composition and location-sharing behavior are inherently selective. Owing to limitations in the demographic attributes available from social media data, information such as age and educational attainment cannot be obtained comprehensively, making it difficult to further identify differences in emotional expression among population groups. The findings therefore primarily reflect the overall emotional expressions of active Weibo users across urban space. Meanwhile, emotional expressions on Weibo may be influenced by salient events, individual experiences, and social media use habits, and their spatial variation cannot be attributed entirely to direct perceptions of the surrounding street environment. Future studies could further incorporate environmental and social factors that may influence emotional expression to provide a more comprehensive analysis of the relationship between SSQ and public sentiment. From a temporal perspective, this study used annual data aggregation to reduce the influence of short-term emotional fluctuations on the overall results; however, this treatment may obscure changes in sentiment across seasons, dates, and specific events. Future research could conduct multi-period analyses to further compare dynamic changes in public sentiment and spatial mismatch patterns.

Second, street-view imagery mainly characterizes the visual environment and cannot fully capture auditory, thermal, tactile, or other sensory experiences. Future studies could develop multisensory SSQ measurement systems to examine street environments and public sentiment from a broader range of dimensions. In addition, street-view imagery is static and cannot represent how people actually use streets across different times and travel paths. Subsequent research could incorporate pedestrian trajectories or mobile-location data to link environmental characteristics with movement patterns and investigate spatial quality and sentiment under dynamic exposure.

Third, the spatial mismatch results may remain sensitive to the modifiable areal unit problem (MAUP), neighborhood definition, and classification criteria. This study used fixed 500 m × 500 m regular grids as the basic analytical units. Different grid resolutions may alter the spatial aggregation of SSQ and public sentiment and the relationships identified between them [90], thereby affecting the resulting mismatch types and their spatial boundaries. Although threshold replacement and Bootstrap resampling were used to test the stability of the same-grid classifications, differences in the results across spatial scales and neighborhood specifications have not yet been systematically examined. Future research could conduct multi-scale sensitivity analyses to further assess the stability of spatial mismatch patterns and combine longitudinal data with multivariate analytical methods to investigate the potential mechanisms underlying their formation.

Finally, this study focused exclusively on the central urban area of Shanghai. Urban characteristics such as high-density development, land-use mix, and street morphology may influence the matching patterns between SSQ and public sentiment. Therefore, the specific mismatch patterns and associated built environment characteristics identified in this study primarily reflect the urban context of central Shanghai and may differ in areas with different stages of urban development and spatial structures. Future research could conduct cross-city comparisons to further examine intercity differences in mismatch patterns and assess the applicability of the multilevel spatial diagnostic framework across different urban contexts.

5. Conclusions

This study integrates street-view imagery and social media data within a multilevel spatial diagnostic framework encompassing same-grid matching, neighborhood spatial association, and classification stability. The framework provides a spatial basis for identifying urban renewal priorities and supporting fine-grained governance and offers a methodological reference for applying this diagnostic framework across different urban contexts. The results indicate that spatial mismatches exist between Street Space Quality and public sentiment in the central urban area of Shanghai and that different mismatch types exhibit distinct spatial characteristics. Although significant neighborhood associations occurred only in some areas, a subset of mismatches displayed directionally consistent neighborhood continuity, suggesting that Street Space Quality–public sentiment mismatch can extend into spatially contiguous areas. In addition, the stable mismatch types exhibited distinct multidimensional Street Space Quality profiles. The comparison of built environment characteristics further showed that the stable types differed in functional configuration, development intensity, road network structure, and the provision of daily service and activity facilities, providing a basis for differentiated street renewal.

These findings suggest a potential disconnect between physical-space improvement and location-based public emotional expression during Shanghai’s transition toward the renewal of existing urban areas. Under conditions of high-density development, high levels of land-use mix, and limited space for new construction, universal additions to the physical environment alone are unlikely to meet the diverse needs of the central urban area. The observed mismatches indicate that places with similar physical foundations may still receive different public responses and that these differences may remain confined to individual street segments or extend into neighboring areas. Urban renewal should therefore move beyond a singular focus on physical quality and toward the coordinated diagnosis of environmental provision and public feedback, enabling limited renewal resources to respond more precisely to local needs. This renewal approach may also provide a useful reference for high-density, mixed-use urban areas facing the need to renew their existing urban fabric.