Measuring urban physical environments using image deep features

Abstract

Effective and efficient representation of the urban physical environment is essential in urban studies. Street-level imagery, a prevalent data source for studying urban physical environments, provides detailed visual information about cityscapes. However, traditional methods typically rely on the audits of predefined semantic elements, which often oversimplify the urban environments. This study introduces a framework that applies image deep features to capture the complexity of urban physical environments, thereby inspiring solutions to urban problem. These deep features encode rich urban landscape semantics into a high-dimensional feature, capturing not only visual elements but also their detailed attributes and complex spatial interrelationships. We validate this approach through a case study on urban landscape cell identification, a task that demands an integrated understanding of urban visual elements and their spatial organization. By analyzing over 8 million street-level images from 36 Chinese cities, we demonstrate that these deep features effectively encode complex patterns and latent structural characteristics of urban environments, surpassing traditional methods. This work offers a novel perspective by introducing image deep features into urban studies, fundamentally enhancing urban understanding and providing a powerful foundation for future research and informed decision-making.