Urban Complexity through Vision Intelligence: Variance, Gradients, and Correlations across Six Italian Cities

TL;DR

The paper tackles the challenge of quantifying urban quality and morphology across six Italian cities using vision-based sensing. It introduces UrbIA Vision Intelligence, deploying 500 Humarels per city to sample Street View imagery and score metrics such as PCI and FDS via GPT-4 Vision prompts, yielding a georeferenced visual census for analyses of Spatial Variance, Urban Gradient, and Cross-Metric correlations. Key findings include pronounced spatial heterogeneity (e.g., Milan $\sigma^2_{PCI} = 1.52$) and consistently weak urban gradients ($R^2 < 0.03$), with a modest positive link between façade quality and greenery ($\rho \approx 0.35$). These results demonstrate the diagnostic potential of vision intelligence for scalable urban analytics and motivate national-scale expansion that can integrate additional contextual data for planning and policy applications.

Abstract

This paper introduces a scalable methodology for the objective analysis of quality metrics across six major Italian metropolitan areas: Rome, Bologna, Florence, Milan, Naples, and Palermo. Leveraging georeferenced Street View imagery and an advanced Urban Vision Intelligence system, we systematically classify the visual environment, focusing on key metrics such as the Pavement Condition Index (PCI) and the Façade Degradation Score (FDS). The findings quantify Structural Heterogeneity (Spatial Variance), revealing significant quality dispersion (e.g., Milan $σ^2_{\mathrm{PCI}}=1.52$), and confirm that the classical Urban Gradient -- quality variation as a function of distance from the core -- is consistently weak across all sampled cities ($R^2 < 0.03$), suggesting a complex, polycentric, and fragmented morphology. In addition, a Cross-Metric Correlation Analysis highlights stable but modest interdependencies among visual dimensions, most notably a consistent positive association between façade quality and greenery ($ρ\approx 0.35$), demonstrating that structural and contextual urban qualities co-vary in weak yet interpretable ways. Together, these results underscore the diagnostic potential of Vision Intelligence for capturing the integrated spatial and morphological structure of Italian cities and motivate a large national-scale analysis.

Figures (5)

• Figure 1: Representative examples of the visual content captured by Humarel agents across the six cities. The top row illustrates structural conditions, including pavement and façade degradation, while the bottom row shows contextual and morphological features such as graffiti, greenery, and the urban canyon effect. Images are included solely for qualitative illustration of the UrbIA vision sensing process.
• Figure 2: Empirical distributions of structural visual quality scores across six Italian metropolitan areas (Humarel sampling, $N{=}500$ per city). Violin plots display the full density of scores (1--5), with medians and extrema. PCI and FDS share the same y-axis range for comparability. The plots complement Table \ref{['tab:structural_quality_stats']} by revealing the internal shape of spatial heterogeneity (e.g., skewness, multimodality), thereby motivating the Urban Gradient analysis introduced in the next section.
• Figure 3: Distribution of the 500 Humarel Sampling Points in the six study cities. The location of the Historical Center (anchor point for DHC) is marked with a distinctive symbol.
• Figure 4: Representative Urban Gradient plots for three Italian metropolitan areas (Milan, Florence, and Palermo). Each scatter plot shows individual Humarel observations (n=500 per city) with fitted regression lines for both Pavement Condition Index (PCI, left column) and Façade Degradation Score (FDS, right column). Despite mild positive or negative slopes, the overall dispersion remains high, confirming that distance from the historical core explains only a small share of the observed spatial variance.
• Figure 5: Spearman correlation heatmap of visual metrics across all cities. Colors range from blue (negative correlation) to red (positive correlation). The strongest relationship appears between Façade and Greenery, while other pairs show weaker, complementary dependencies.