Measuring human perceptions of streetscapes to better inform urban renewal: A perspective of scene semantic parsing

Ubiquitous and up-to-date geotagged data are increasingly employed to uncover the visual traits of the built environment. However, few prior studies currently link this theoretical knowledge of street appraisals with operable practices to inform streetscape transformation. This study proposes a proof-of-concept analytical framework that sheds light on the connections between urban renewal and the quantification of streetscape visual traits. By virtue of a million intensively collected panoramic street view images in Shenzhen, China, the image-segmentation technique SegNet automatically extracts pixelwise semantical information and classifies visual elements. The throughput of the eye-level perception of the street canyon is formed by five indices. Additionally, the framework-derived scores (FDSs) are contrasted with the subjective rating scores (SRSs) to report the divergence and coherence between the visually experienced and the quantitative estimated methods. Furthermore, we investigate the spatial heterogeneity of five perception aspects, discuss the variations of the perception outcomes across different function streets, and analyze the net effect of urban renewal projects (URPs) on streetscape transformation. We conclude that this deep learning-driven approach provides a feasible paradigm to depict high-resolution streetscape perception, to analyze fine-scale built environment, and to effectively bridge gaps between the street semantic metric and urban renewal.