Cities are increasingly viewed as complex systems with interdependent social, economic, and environmental components. Urban street networks are crucial for understanding these systems, and their analysis has been applied to various urban phenomena, including spatial homogeneity, human development, and traffic forecasting. However, standard network analysis methods often overlook the rich diversity of city landscapes and the contextual factors influencing network structures and flows. Existing solutions like Knowledge Graphs and Digital Twins pose significant barriers to adoption due to resource demands. To address these challenges, Urbanity offers an automated, scalable, and lightweight approach to building detailed urban networks. It integrates diverse city indicators (building shapes, street views, population counts, points of interest) into network representations, supporting various analysis and modeling methods. The goal is to facilitate easier and more comprehensive study of complex urban systems, enabling automated network construction at any scale and providing summary indicators for any geographic area.

The paper reviews existing tools for urban network analysis, categorizing them based on functionality (primal/dual network representations, wayfinding, urban accessibility, clustering). Most tools focus primarily on distance-based and topological indicators, with limited integration of contextual information. Urbanity extends the existing ecosystem by improving the user interface, enhancing network feature representation (incorporating context-based and semantic indicators), and providing efficient benchmarking functionalities. This allows for applications such as multi-criteria site analysis, graph predictive modeling, and visualization of networks. The authors highlight the limitations of existing tools focusing solely on distance and connectivity metrics, which fail to capture the complexity of urban systems. They cite research suggesting the use of more detailed deep learning models with location-specific features to address this limitation.

Urbanity is built using Python, leveraging existing packages for network analysis and data processing. Data acquisition involves utilizing publicly available datasets with global coverage and fine spatial resolution, including OpenStreetMap for street networks, building footprints, and points of interest; Mapillary for street view imagery; and Meta's high-resolution population density maps. The package provides a high-level interface for data extraction and preprocessing. The core of Urbanity is built around network representation using the Networkx package, allowing for graph-based analysis. The package automates the computation of various contextual indicators and provides functionality for computing aggregate statistics for any geographic area of interest. An edge classification task was implemented using graph neural networks to predict road categories (national, regional, precinct, neighborhood, local access) in the five study cities. Various architectures (GCN, GAT, GraphSAGE) were compared using ablation studies to evaluate feature importance (metric/topological, building, population, points of interest, street view imagery). Model performance was assessed using mean classification accuracy. Image segmentation for street view imagery indicators was performed using the Mask2Former model.

The Urbanity global network dataset reveals significant differences in urban network structures and indicators across the five study cities (Singapore, Paris, Bangkok, Chicago, Seattle). Bangkok exhibited the highest population count and lowest building footprint percentage, suggesting urban sprawl. Paris, Chicago, and Seattle displayed higher network density than Bangkok. Singapore, a city-state, stood out with high population density despite lower building footprint and network density compared to others. Network assortativity analysis of green view index (GVI) in Singapore and Bangkok showed that Singapore demonstrated well-balanced urban greenery, while Bangkok showed segregation patterns consistent with existing green zone regulations. Analysis comparing node density and mean building complexity across cities and subzones revealed distinct patterns reflecting each city's planning history and context. Paris exhibited high homogeneity, while Singapore showed high diversity, indicating varied land use requirements. Bangkok, Chicago, and Seattle showed a positive correlation between these two indicators. The edge classification experiments showed that incorporating semantic and contextual indicators consistently improved predictive accuracy of road categories across all cities. Graph Attention Networks (GAT) generally outperformed other GCN architectures. The removal of metric and topological features significantly reduced performance, highlighting the importance of graph structure information. Optimal parameters (e.g., node buffer radius) varied across cities, emphasizing the need for context-specific model tuning.

The findings highlight the value of a comprehensive approach to urban network analysis that integrates contextual features. The study contributes to understanding urban complexity by incorporating diverse indicators beyond traditional network metrics and topological properties. The results demonstrate the potential of Urbanity for various applications in urban planning and analysis, including multi-criteria site selection, predictive modeling, and comparative urban studies. The use of open-source tools and data promotes open science and facilitates broader access to advanced urban analytics. The importance of balancing computational efficiency with equitable, inclusive planning practices is emphasized. Future work should focus on enriching the network representation with finer-grained data (e.g., sidewalk conditions) and incorporating dynamic modeling techniques (e.g., reinforcement learning) to capture the evolution of urban systems.

Urbanity, an open-source Python package, automates the construction of feature-rich urban networks, enhancing urban analysis capabilities. The study demonstrated the significant contribution of contextual features to improving the accuracy of descriptive and predictive models for understanding urban systems. Future development should focus on integrating more detailed contextual information, improving the temporal resolution of data, and incorporating agent-based modeling to capture the dynamics of urban systems. The open-source nature of Urbanity fosters collaboration and broader access to advanced urban analytics tools.

While Urbanity addresses several limitations of existing tools, some limitations remain. Data availability and quality can vary across cities and datasets. The performance of machine learning models depends on the quality and representativeness of the training data, and model parameters may require tuning for specific urban contexts. Furthermore, the current version of Urbanity primarily focuses on street networks, and future development should include other types of urban networks. The study is limited to five cities, and further research with a wider range of cities is needed to generalize the findings.