Quantifying physical and psychological perceptions of urban scenes using deep learning

Abstract

The complicated relationship between urban scenes and public perceptions has long been a concern in many disciplines. Previous studies have lacked human-oriented technical paths and high-throughput datasets to quantify physical and psychological perceptions in different land-use scenarios. This paper adopts a novel transfer learning approach to quantify the six types of landsense indices (LSIs) as psychological perception metrics and employs panoptic segmentation to parameterize the view index (VI) and the number of foreground instances (NFIs) as physical perception measures. Then, a quantitative analysis is conducted in Beijing’s six Ring Road areas, and the connections between people’s physical and psychological perceptions of heterogeneous land use are explored. The landsense maps can depict the distribution of LSIs and facilitate the understanding of complex perceptions distributed at a large scale. The regression model shows that natural landscapes (trees, grasses, and mountains) in the Beijing built-up area exhibit an overall positive performance. Moreover, for several block-level land uses, industrial scenery is related to overall negative psychological feelings. Parks and green spaces are positively related to psychological perceptions, because of the greater exposure opportunities to natural landscapes for residents. The framework in this research has potential in assisting urban planning and land-use management, and it enriches the datasets with extensive information, thereby improving the psychological perceptions of urban scenes from residents’ perspectives. The novel approaches in this paper take a step forward in quantifying and understanding the public perceptions of urban landscapes.