Comprehensive street built environmental recognizabililty evaluation by integrating visual and spatial structural data

Abstract

Evaluating the recognizability of street built environments provides crucial support for urban planning, security monitoring and navigation. Although street view images (SVIs) are widely used in urban studies, it overlooks the interconnection among different locations, which can also affect perceptions about environmental recognizability. To address this issue, this study proposes a deep learning-based model called RB-Node, which comprehensively integrates spatial structural features in a road network view and visual features from SVIs, achieving 82.56% accuracy. It appears that image information from visual features dominates environmental recognizability. Additionally, structural information contributes significantly to the accurate classification of nodes and waterfront promenade areas. Moreover, scene-text information, a subset of visual features, helps classify commercial and historical areas. Furthermore, 1056 samples were collected through an eye-tracking experiment to validate the recognizability evaluation results, as well as compare the decision-making process between humans and RB-Node. According to the results, RB-Node behaviour and human observed behavior follow similar patterns, although human perceptions tend to be more holistic than RB-Node's. This study contributes to a better understanding of environmental recognizability and provides targeted recommendations for urban renewal.

Keywords

• Environmental recognizability evaluation
• Multi-source data
• Street view image
• Road network
• Deep learning
• Street built environment