Impacts of Micro-Scale Built Environment Features on Residential Location Choice: a computer vision-aided assessment

Abstract

Recent studies have highlighted the significant impact of built environments (BE) in residential neighbourhoods on well-being, focusing on correlations between micro-scale BE features—such as trees, grass, fences, and bikes—and specific well-being aspects like physical health, social interaction, or perceptions of safety. However, these studies often fail to comprehensively explore how these features influence residential preferences or provide clear design guidance. To address this gap, examining residential location choice (RLC) datasets can offer valuable insights into how various micro-scale BE features affect the attractiveness of residential neighbourhoods and residents' well-being. This thesis proposes a semantic Computer Vision-enriched Discrete Choice Model (CV-DCM), which uses a panoptic segmentation model to quantify micro-scale BE features and integrates with traditional discrete choice models.

A manual evaluation of 400 masks generated by the semantic computer vision model ensured accurate quantification of BE features, enhancing choice modelling interpretation. The choice modelling results highlight the specific impacts of various micro-scale BE features on RLC, providing valuable insights for urban planners. For instance, restricting unattractive features like motorcycles in residential neighbourhoods and planting or maintaining more trees are the most attractive among vegetation. By pinpointing the most influential elements, the study facilitates the cost-effective restructuring of residential neighbourhoods to boost their attractiveness and enhance residents' well-being.