Objective This research aims to systematically investigate the spatial heterogeneity in the driving mechanisms behind human perception of street visual vitality, with the broader goal of supporting more refined and human-centered urban street planning and management. Street visual vitality, as a perception-based indicator, reflects the degree to which people feel a street is active, engaging, and comfortable. Unlike traditional measures that rely mainly on land use or traffic data, perception-driven approaches capture how people actually experience and evaluate the urban environment. By examining how different visual elements, such as the proportion of vegetation, the density of buildings, the presence of signage, micro-mobility activities, and the openness of the sky, affect vitality perception across varying spatial contexts, this research seeks to reveal both general patterns and local nuances.
Methods This research adopts a comprehensive, multi-stage analytical framework that integrates deep learning with spatially adaptive statistical modeling, focusing on the six central districts of Tianjin, China. First, a large-scale longitudinal dataset of street view images from 2013 to 2020 was established through automated web scraping. Based on this dataset, a ResNet50 deep learning model was trained using perception-labeled samples to estimate the visual vitality score of each street view image. The model was trained to recognize subtle environmental cues, such as human presence, facade articulation, greenery coverage, and traffic context, that jointly contribute to human vitality perception, enabling consistent and reliable predictions across the eight-year period. Second, to extract structured streetscape information, the DeepLab V3+ semantic segmentation model was applied. Through this process, the originally unstructured pixel data was transformed into interpretable visual features that represent real physical components of the street environment. To reduce the dimensionality and complexity of the large feature set while preserving spatial differences, the study applied Geographically Weighted Principal Component Analysis. Unlike traditional principal component analysis, which produces global components, this method identifies localized combinations of visual features that may vary across regions. This helps capture the fact that similar visual attributes can reflect different environmental meanings depending on where they appear in the city. Finally, to explore how these features influence visual vitality perception in a spatially heterogeneous way, a Geographically Weighted Random Forest model was employed. This model combines the nonlinear learning ability of random forests with a spatial weighting mechanism, allowing each location to have its own model structure and variable importance ranking. This approach makes it possible to detect how the same visual features may have stronger, weaker, or even reversed effects in different parts of the city.
Results The empirical analysis reveals three major findings: 1) The spatial distribution of visual vitality perception exhibits a stable pattern characterized by higher values in the central districts and lower values in the peripheral districts. Over the eight-year period, vitality in the central districts gradually increased, likely due to continuous improvements in public space quality, transport infrastructure, and service facilities. In contrast, vitality in the peripheral districts experienced small fluctuations and a slight downward trend, reflecting relatively slower development or fewer public-realm enhancements. This center-periphery contrast highlights the uneven progression of urban vitality across the city. 2) While vitality itself is highest in the center, the influence intensity of the six principal components shows the opposite pattern: the central areas display lower influence intensity, whereas the peripheral areas show higher intensity. This suggests that central districts possess more complex and diverse environmental features, causing no single factor to dominate the perception outcome. In contrast, peripheral districts often rely on a smaller set of environmental characteristics, such as boundary elements, large transport interfaces, or micro-mobility activities, which exert stronger and more concentrated effects on visual vitality perception. The analysis also identifies a directional difference across the study area, with northeastern districts generally showing higher perceived vitality and southwestern districts consistently showing lower vitality, reflecting broader socio-economic and spatial development patterns. 3) The dominant driving components vary across spatial contexts. In the central districts, PC6 consistently emerges as the most influential driver, indicating that features related to traffic signs, poles, and active mobility contribute significantly to visual vitality perception in more developed areas. In peripheral districts, however, the influence is more diverse: PC1 and PC3 interact and jointly shape the perception results. This reveals that street vitality in less developed areas is affected by a complex mix of boundary characteristics, greenery structure, building configurations, and transportation elements. Such findings underscore the need to design targeted improvement strategies that address the specific environmental conditions of each area.
Conclusion The results demonstrate that street vitality is not driven by a single universal factor but by a set of features whose importance varies across space and time. These findings offer valuable insights for urban planners, designers, and policymakers. In central districts, enhancing street vitality may involve improving multimodal transportation environments, refining street signage and facade quality, and optimizing pedestrian circulation. In peripheral districts, more substantial benefits may be achieved by enhancing boundary permeability, enriching greenery and public space elements, and strengthening micro-mobility connections. Overall, these findings highlight the importance of considering spatial heterogeneity in urban planning and demonstrate how advanced computational techniques can support more responsive, equitable, and human-centered urban design decisions.
DownLoad: