Diagnosing Urban Street Vitality via a Visual-Semantic and Spatiotemporal Framework for Street-Level Economics

Xinxin Zhuo; Mengyuan Niu; Ruizhe Wang; Junyan Yang; Qiao Wang

cs.CYApr 10, 2026

Diagnosing Urban Street Vitality via a Visual-Semantic and Spatiotemporal Framework for Street-Level Economics

Authors: Xinxin Zhuo, Mengyuan Niu, Ruizhe Wang, Junyan Yang, Qiao Wang

Organizations: School of Economics and Management, Southeast University, Nanjing, China · School of Information Science and Engineering, Southeast University, Nanjing, China · School of Architecture, Southeast University, Nanjing, China

Abstract

Micro-scale street-level economic assessment is fundamental for precision spatial resource allocation. While Street View Imagery (SVI) advances urban sensing, existing approaches remain semantically superficial and overlook brand hierarchy heterogeneity and structural recession. To address this, we propose a visual-semantic and field-based spatiotemporal framework, operationalized via the Street Economic Vitality Index (SEVI). Our approach integrates physical and semantic streetscape parsing through instance segmentation of signboards, glass interfaces, and storefront closures. A dual-stage VLM-LLM pipeline standardizes signage into global hierarchies to quantify a spatially smoothed brand premium index. To overcome static SVI limitations, we introduce a temporal lag design using Location-Based Services (LBS) data to capture realized demand. Combined with a category-weighted Gaussian spillover model, we construct a three-dimensional diagnostic system covering Commercial Activity, Spatial Utilization, and Physical Environment. Experiments based on time-lagged geographically weighted regression across eight tidal periods in Nanjing reveal quasi-causal spatiotemporal heterogeneity. Street vibrancy arises from interactions between hierarchical brand clustering and mall-induced externalities. High-quality interfaces show peak attraction during midday and evening, while structural recession produces a lagged nighttime repulsion effect. The framework offers evidence-based support for precision spatial governance.

Figures & tables

Figure 1. Study area. A map showing the central metropolitan area of Nanjing, Jiangsu Province, which serves as the study area.

Figure 2. Analysis framework. A flowchart illustrating the multimodal data fusion pipeline, from SVI acquisition to visual-semantic interpretation.

Figure 3. Spatiotemporal evolution of dynamic crowd intensity (UV) across four tidal periods for both a weekday and a weekend in Nanjing. A 2x4 matrix of heatmaps showing the dynamic crowd intensity across different tidal periods. The hotspots shift and expand significantly during the midday and night economy periods compared to the morning peak.

Figure 4. Training performance of YOLOv5-seg. Line charts displaying the loss and mAP training curves over 100 epochs for the YOLOv5-seg model.

Figure 5. The architecture of the VLM-LLM Brand Decoding Pipeline. A diagram showing the two-stage VLM-LLM pipeline for extracting and refining brand semantics from storefront images.

Brand Type	GT Count	Metric	OCR	VLM Only	VLM + LLM (Ours)
International	99	Precision	0.474	0.638	0.802
		Recall	0.374	0.606	0.737
		$F1$ -score	0.418	0.622	0.768
Ordinary	54	Precision	0.449	0.760	1.000
		Recall	0.407	0.704	0.852
		$F1$ -score	0.427	0.731	0.920

Table 1. Performance comparison of brand recognition pipelines (Ablation Study)

Category	Visual Example	Description
Signboard		A proxy for commercial density and the primary carrier of semantic brand information, reflecting the qualitative hierarchy of local businesses.
Closed Store		A visual indicator of economic attrition; captures vacancy dynamics and hidden recession often missed by static POI datasets.
Glass Interface		Represents the permeability of the street canyon, facilitating visual interaction between indoor commerce and outdoor pedestrians.
Person		Direct proxy for social vibrancy; reflects the street’s attractiveness and capacity for social interaction.
Bicycle		Reflects sustainable mobility patterns and green accessibility.
Car		Indicates vehicular accessibility and commercial logistics intensity.

Table 2. Annotated visual categories and their renewal-oriented interpretations.

Index	Value
Hardware Environment	5 $\times$ NVIDIA GeForce RTX 3090 (Data Parallel)
Batch size	8
Epochs	100
Training images	642
Learning rate	0.01

Table 3. Training Parameters

Variable	Formula	Definition of Components
Shop Density	$SD_{i}=\frac{NS_{L_{i}}+NS_{R_{i}}}{l_{i}}$	$NS$ : Number of signboards detected on the respective side.
Closure Ratio	$CR_{i}=\frac{NC_{L_{i}}+NC_{R_{i}}}{NS_{L_{i}}+NS_{R_{i}}}$	$NC$ : Number of closed stores (identified by roller shutters/locked gates). $NS$ : Total count of signboards (active + closed).
Weighted Brand Ratio	$BR_{i}=\frac{NB^{w}_{L_{i}}+NB^{w}_{R_{i}}}{NS_{L_{i}}+NS_{R_{i}}}$	$NB^{w}$ : Weighted brand score (numerator). Calculation: $NB^{w}=N_{loc}\cdot w_{1}+N_{glb}\cdot w_{2}$ $N_{loc/glb}$ : Number of Local/International brands. $w_{1},w_{2}$ : Hierarchy weights (e.g., $1.0$ vs $1.5$ ). Note: Using the total signboard count (NS) as the denominator inherently penalizes streets with high closure rates, reflecting the dilution of brand premium in decaying corridors.
Mall Spillover Vitality	$MV_{i}=\sum_{j=1}^{N}\mathbb{I}(d_{ij}\leq D)\cdot\exp\left(-\frac{d_{ij}^{2}}{2\sigma_{j}^{2}}\right)$	$d_{ij}$ : Euclidean distance between street unit $i$ and mall $j$ . $\sigma_{j}$ : Gaussian bandwidth specific to mall $j$ ’s hierarchy. $\mathbb{I}(\cdot)$ : Indicator function (1 if $d_{ij}\leq D$ , 0 otherwise). $D$ : Max spatial threshold (e.g., 2000m).
Motor Vehicle Density	$MD_{i}=\frac{NM_{L_{i}}+NM_{R_{i}}}{l_{i}}$	$NM$ : Number of detected motor vehicles (cars, trucks, buses).
Non-Motor Vehicle Density	$ND_{i}=\frac{NN_{L_{i}}+NN_{R_{i}}}{l_{i}}$	$NN$ : Number of non-motor vehicles (bicycles, e-scooters).

Table 4. Summary of variables, calculation formulas, and definitions.

Figure 6. The distribution of variables at the sampling-point level. A series of maps showing the spatial distribution of all nine vitality variables across the sampling points.

Figure 7. The distribution of Commercial Activity at the sampling-point level. A spatial distribution map illustrating the concentration of Commercial Activity indicators at the sampling-point level.

Figure 8. The distribution of Spatial Utilization at the sampling-point level. A spatial distribution map illustrating the concentration of Spatial Utilization indicators at the sampling-point level.

Figure 9. The distribution of Physical Environment at the sampling-point level. A spatial distribution map illustrating the concentration of Physical Environment indicators at the sampling-point level.

Figure 10. Spearman correlation heatmap identifying the internal coupling of vitality indicators. A Spearman correlation heatmap showing the relationships among the nine vitality indicators.

Figure 11. Principal Component Analysis results. A chart visualizing the Principal Component Analysis (PCA) results and factor loadings.

Variables	PC1	PC2	PC3	PC4
Pedestrian Presence	0.411	0.037	-0.094	0.110
Motor Vehicle Density	0.420	-0.019	-0.044	-0.088
Non-Motor Vehicle Density	0.230	0.328	-0.164	0.800
Green Coverage Ratio	0.149	0.745	-0.081	-0.519
Weighted Brand Ratio	0.383	-0.141	0.007	-0.179
Shop Density	0.437	-0.290	0.026	-0.140

Table 5. Principal component loading matrix of key indicators (PC1 to PC4)

Figure 12. The distribution of SEVI. A spatial distribution map showing the Street Economic Vitality Index (SEVI) across the study area.

Figure 13. Spatiotemporal evolution of the Time-Sliced GWR model’s explanatory power (Adjusted $R^{2}$ ). A line graph showing the Adjusted R-squared values for the GWR model across eight time periods, highlighting peaks during midday and evening.

Figure 14. Temporal evolution of local GWR coefficients for Mall Spillover Vitality and Closure Ratio. Two boxplots showing the distribution of local GWR coefficients across eight time periods. The top plot shows Mall Spillover Effect, and the bottom shows the negative impact of Closed Shops.

Time Period	1000m	2000m (Baseline)	3000m
Weekday (10.15) Morning	0.5745	0.5729	0.5743
Weekday (10.15) Midday	0.7125	0.7089	0.7164
Weekday (10.15) Evening	0.6840	0.6800	0.6808
Weekday (10.15) Night	0.6609	0.6629	0.6620
Weekend (10.19) Morning	0.5945	0.5974	0.5942
Weekend (10.19) Midday	0.7006	0.6985	0.7030

Table A1. Comparison of Time-Lagged GWR Model Explanatory Power (Adjusted

R^{2}

) under Varying Spatial Thresholds

Figure A1. Temporal evolution of local GWR coefficients for Mall Spillover Vitality under alternative spatial thresholds (1000m, 2000m, and 3000m). The consistent positive medians and expanding interquartile ranges during non-commuting hours confirm the structural robustness of the field-based spillover effect.

Time Period	Gaussian Decay (Baseline)	Exponential Decay	Linear Decay
Weekday (10.15) Morning	0.5729	0.5734	0.5738
Weekday (10.15) Midday	0.7089	0.7164	0.7121
Weekday (10.15) Evening	0.6800	0.6808	0.6838
Weekday (10.15) Night	0.6629	0.6623	0.6695
Weekend (10.19) Morning	0.5974	0.5970	0.5973
Weekend (10.19) Midday	0.6985	0.7041	0.7049

Table A2. Comparison of Time-Lagged GWR Model Explanatory Power (Adjusted

R^{2}

) under Alternative Distance-Decay Functions

Indices	SEVI EWM-TOPSIS	SEVI EQ	SEVI PCA
SEVI EWM-TOPSIS	1.000	0.994 ∗∗∗	0.977 ∗∗∗
SEVI EQ	0.994 ∗∗∗	1.000	0.984 ∗∗∗
SEVI PCA	0.977 ∗∗∗	0.984 ∗∗∗	1.000

Table A3. Spearman Rank Correlation Matrix of Alternative SEVI Constructs

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

$BR_{i}$ Intensity Tier	Mean Total POI Count	Mean Premium/Discretionary POI Count
Low-tier Brand Segments	7.58	0.93
Mid-tier Brand Segments	10.00	1.24
High-tier Brand Segments	12.94	1.61
Growth (High vs. Low)	+70.7%	+73.1%

Appendix

Table A4. External Validation of

BR_{i}

against Physical Business Density

Urban Planning

Urban planning uses geospatial data, machine learning, and optimization to analyze cities and guide transport, land use, housing, and climate decisions.

See all 47 papers

Bridging Geographic Bias in Urban Streetscape Inference via Lifelong Learning with Visual-Semantic Pivoting
Jun 13, 2026Xinze ZhangGeospatial Representation Learning Continual Learning
University of Southern California, Los Angeles, CA 90007, USA
From Street View Imagery to Street Quality Indicators: Vision Language Inference for the Suburban 15-minute City
Aug 20, 2026Joan Perez, Giovanni FuscoUrban Planning
Urban Geo Analytics, France · Université Côte-Azur-CNRS-AMU-Avignon Université, ESPACE, France
Cross-View Urban Sensing: Mapping Subjective Streetscape Perception via AlphaEarth Embeddings and Urban Context
Aug 17, 2026Peilin Li, Pengfei Chen, Jingyu Wang +4Urban Planning
School of Geospatial Engineering and Science, Sun Yat-sen University, and Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai, 519082, China · Key Laboratory of Comprehensive Observation of Polar Environment (Sun Yat-sen University), Ministry of Education, Zhuhai, 519082, China

Jun 13, 2026J/K move · Enter open · S save

Bridging Geographic Bias in Urban Streetscape Inference via Lifelong Learning with Visual-Semantic Pivoting

Xinze Zhang

University of Southern California, Los Angeles, CA 90007, USA

Geospatial Representation Learning Continual Learning Hierarchical Representation Learning Algorithmic Bias

Open paper AskPDF arXiv

Visual perception of urban streetscapes underpins evidence-based decisions in landscape planning, public health, and place-making. Yet models trained on a few well-photographed metropolises systematically misjudge underrepresented districts, propagating geographic bias into downstream policy. We address this gap with HVSP-LL, a lifelong learning framework that couples a stratified visual-semantic pivoting module with an equity-aware rehearsal mechanism. The pivoting module organises landscape concepts along a three-tier ontology (macro structure, meso composition, micro element) and aligns image features to learnable semantic anchors at each tier, providing transferable representations that resist distributional drift. The lifelong adaptation component sequentially absorbs new urban regions while constraining inter-region perception gaps through a worst-region sample-reweighting objective and a structurally-aware exemplar buffer. We evaluate HVSP-LL on a panoramic streetscape benchmark assembled from twelve cities across four continents and seven perceptual dimensions. The framework attains 0.834 Spearman correlation on the held-out city sequence, an absolute 6.1 point improvement over the strongest continual baseline, and shrinks the inter-city perception gap to 0.094 -- a 38% reduction relative to the strongest continual baseline (0.151) and a 57% reduction relative to a representative regularisation baseline (0.218). Ablations confirm that each tier of the pivoting hierarchy contributes monotonically, and the equity-aware rehearsal converts mean backward transfer from -0.038 (without retention) to +0.013, eliminating catastrophic forgetting on the held-out sequence. Our results indicate that hierarchical anchoring is a practical pathway toward geographically equitable streetscape inference at city scale.

Abstract

Urban Planning

Explore similar work

Bridging Geographic Bias in Urban Streetscape Inference via Lifelong Learning with Visual-Semantic Pivoting

From Street View Imagery to Street Quality Indicators: Vision Language Inference for the Suburban 15-minute City

Cross-View Urban Sensing: Mapping Subjective Streetscape Perception via AlphaEarth Embeddings and Urban Context