Restorative perception of urban streets: Interpretation using deep learning and MGWR models

• Published in: Frontiers in public health 2023-03, Vol.11, p.1141630-1141630
• Main Authors: Han, Xin, Wang, Lei, He, Jie, Jung, Taeyeol
• Format: Article
• Language: English
• Subjects: Cities; Deep Learning; Environment; Humans; multiscale geographically weighted regression; Perception; Public Health; restorative quality; semantic segmentation; Spatial Regression; street view; urban street environment
• ISSN: 2296-2565; 2296-2565
• DOI: 10.3389/fpubh.2023.1141630

Restorative environments help people recover from mental fatigue and negative emotional and physical reactions to stress. Excellent restorative environments in urban streets help people focus and improve their daily behavioral performance, allowing them to regain efficient information processing skills and cognitive levels. High-density urban spaces create obstacles in resident interactions with the natural environment. For urban residents, the restorative function of the urban space is more important than that of the natural environment in the suburbs. An urban street is a spatial carrier used by residents on a daily basis; thus, the urban street has considerable practical value in terms of improving the urban environment to have effective restorative function. Thus, in this study, we explored a method to determine the perceived restorability of urban streets using street view data, deep learning models, and the Ordinary Least Squares (OLS), the multiscale geographically weighted regression (MGWR) model. We performed an empirical study in the Nanshan District of Shenzhen, China. Nanshan District is a typical high-density city area in China with a large population and limited urban resources. Using the street view images of the study area, a deep learning scoring model was developed, the SegNet algorithm was introduced to segment and classify the visual street elements, and a random forest algorithm based on the restorative factor scale was employed to evaluate the restorative perception of urban streets. In this study, spatial heterogeneity could be observed in the restorative perception data, and the MGWR models yielded higher interpretation strength in terms of processing the urban street restorative data compared to the ordinary least squares and geographically weighted regression (GWR) models. The MGWR model is a regression model that uses different bandwidths for different visual street elements, thereby allowing additional detailed observation of the extent and relevance of the impact of different elements on restorative perception. Our research also supports the exploration of the size of areas where heterogeneity exists in space for each visual street element. We believe that our results can help develop informed design guidelines to enhance street restorative and help professionals develop targeted design improvement concepts based on the restorative nature of the urban street.