Relationships between cityscape patterns and culture groups revealed by a global genealogy

Yu Ye, Xinghan Chen, Xingchen Zeng, Ling Li, Yuxuan Liu, Y.-s. Liu, Xiangning Li, Lan Wang, Xing Shi, Wei ZengResearch Square 2025 · DOI

Abstract

The absence of large-scale cityscape measurements hampers further exploration of the relationship between cityscapes and cultures, a key focus in cultural geography. As a response, a data-informed analytical method has been developed to enable the measurement of cityscapes at a global scale. Specifically, six key cityscape features were identified and measured using a Large Language Model and deep learning algorithms. Computed cityscape patterns via neural embeddings and UMAP across ten cultural groups were systematically identified for the first time. A significant correlation was found between cityscapes and cultural group classifications, stronger than natural environmental factors. The First Law of Geography holds within a 5,000 km range, beyond which cityscape similarity and geographical distance become largely independent. Long-distance similarities arise from shared configurations shaped by colonial legacies and modernist planning, and are better explained by the Third Law. By integrating intelligent algorithms and big data, this study offers a big picture of global cityscapes and provides new insights into core debates in cultural geography. It aims to inspire a paradigm shift and establish a new frontier at the intersection of cultural geography and computational urban science.

Introduction

For decades, cultural geographers have debated whether nature or human choice primarily shapes the "look and feel" of a city. This study, representing a significant leap in computational urban science, utilizes computer vision and Large Language Models (LLMs) to analyze 6 million street-level images across 90 global cities. By creating a "global genealogy" of cityscapes, the researchers provide quantitative evidence that culture—not just geography or climate—is the primary architect of our urban environments.

Methodology

The authors developed a multi-stage pipeline to transform qualitative urban characteristics into quantitative data. First, GPT-4 was used to identify six key features defining a cityscape: street scale, building color, façade material, architectural style, greenery, and urban signs. Next, deep learning models (including ResNet-50 and Vision Transformers) extracted these features from a massive dataset of Street View Images (SVIs).

The data-informed analytical pipeline used to identify features, extract data, and construct the global cityscape genealogy.

To treat the cityscape as a holistic concept, a Multilayer Perceptron (MLP) fused these features into high-dimensional "latent embeddings." Finally, UMAP (Uniform Manifold Approximation and Projection) was employed to compress these embeddings into 2D maps, revealing how cities cluster visually based on cultural frameworks like the GLOBE categories.

Key Insights and Results

Culture vs. Nature

A primary finding is that cultural group classification is a significantly stronger predictor of cityscape similarity than environmental factors such as climate, elevation, or topography. This suggests that human "genius loci" (the spirit of a place) exerts more influence on urban form than the physical terrain.

The 5,000 km Threshold

The study tested Tobler’s First Law of Geography (proximity breeds similarity) and found it holds true only up to 5,000 km. Beyond this distance, geographical proximity becomes irrelevant to visual similarity. Instead, long-distance similarities are better explained by shared "configurations"—colonial legacies (e.g., British influence on Vancouver and Melbourne) or the global spread of modernist planning.

UMAP visualizations showing the clustering of cityscape patterns and specific city groups based on cultural identity.

Internal Consistency and Homogenization

Different cultural groups exhibit varying levels of visual cohesion. Confucian Asia showed the highest internal similarity, while the Anglo group was the most heterogeneous, reflecting diverse urban evolutions across the US, UK, and Australia. While some regions show signs of visual homogenization due to rapid modern development, most cultural groups maintain distinct, differentiated cityscape patterns.

Conclusion

By integrating intelligent algorithms and big data, this research offers a new perspective on global urban form. The AI’s "visual sense" achieved 83% agreement with human experts, validating the method's accuracy. Future work will aim to include smaller cities and rural areas to determine if cultural "fingerprints" are even more pronounced outside of globalized megacities, and to explore how these spaces are functionally used by their inhabitants.

Summary updated

Abstract

<title>Abstract</title> The absence of large-scale cityscape measurements hampers further exploration of the relationship between cityscapes and cultures, a key focus in cultural geography. As a response, a data-informed analytical method has been developed to enable the measurement of cityscapes at a global scale. Specifically, six key cityscape features were identified and measured using a Large Language Model and deep learning algorithms. Computed cityscape patterns via neural embeddings and UMAP across ten cultural groups were systematically identified for the first time. A significant correlation was found between cityscapes and cultural group classifications, stronger than natural environmental factors. The First Law of Geography holds within a 5,000 km range, beyond which cityscape similarity and geographical distance become largely independent. Long-distance similarities arise from shared configurations shaped by colonial legacies and modernist planning, and are better explained by the Third Law. By integrating intelligent algorithms and big data, this study offers a big picture of global cityscapes and provides new insights into core debates in cultural geography. It aims to inspire a paradigm shift and establish a new frontier at the intersection of cultural geography and computational urban science.
BibTeX
@inproceedings{ye2025relationships,
  author = {Yu Ye and Xinghan Chen and Xingchen Zeng and Ling Li and Yuxuan Liu and Y.-S. Liu and Xiangning Li and Lan Wang and Xing Shi and Wei Zeng},
  title = {{Relationships between cityscape patterns and culture groups revealed by a global genealogy}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-5047637/v1},
}
Figure