Characterising urban neighbourhoods using street view images: unsupervised clustering and perception-based mapping
File(s)
Author(s)
Muller, Emily
Type
Thesis
Abstract
The way cities are built shape our interaction with urban environments through both physical and social dimensions that can have lasting impact on the well-being of urban populations. Understanding these impacts requires effective measurement of upstream determinants. While survey techniques offer a direct means of measurement, their implementation is often hindered by cost and scalability, particularly when aiming for city-wide coverage. Conversely, routinely collected administrative data, while scalable, may lack specificity in capturing desired features.
This thesis adopts a street-view image-based approach to area-level urban measurement assuming scarcity of outcome labels. The reason for this assumption is two-fold; in the absence of outcome labels, the framework proposed is transferable to data scarce urban settings, and secondly, this aims to circumvent overfitting and therefore offer greater generalisability.
Measurement is achieved using a two-pronged approach, one which characterises the physical dimensions of the urban environment, and another to characterise the social. Physical characterisation is achieved using unsupervised clustering with a Human-In-The-Loop interpretability framework. Unsupervised clustering returns urban classes related to land uses such as commercial zones, different types of residential areas, and green spaces. The subjective dimension is addressed through a computational framework that learns and predicts perceptual scores related to walkability, safety, aesthetics, socioeconomic status, and emotional impressions, aggregated at the neighbourhood level.
Returning to urban health of residents, the relationship between the newly created area-level measurements and determinants of health are examined and found to explain variance in pollution estimates and life expectancy.
With the promise of street view imagery, ethical considerations and potential biases warrant scrutiny. Ongoing research must assess its efficacy in real-world urban health applications, identifying promise and limitations. In sum, this thesis contributes methodological approaches to urban characterisation using street-level imagery, offering a reproducible framework for urban health researchers to leverage in their investigations.
This thesis adopts a street-view image-based approach to area-level urban measurement assuming scarcity of outcome labels. The reason for this assumption is two-fold; in the absence of outcome labels, the framework proposed is transferable to data scarce urban settings, and secondly, this aims to circumvent overfitting and therefore offer greater generalisability.
Measurement is achieved using a two-pronged approach, one which characterises the physical dimensions of the urban environment, and another to characterise the social. Physical characterisation is achieved using unsupervised clustering with a Human-In-The-Loop interpretability framework. Unsupervised clustering returns urban classes related to land uses such as commercial zones, different types of residential areas, and green spaces. The subjective dimension is addressed through a computational framework that learns and predicts perceptual scores related to walkability, safety, aesthetics, socioeconomic status, and emotional impressions, aggregated at the neighbourhood level.
Returning to urban health of residents, the relationship between the newly created area-level measurements and determinants of health are examined and found to explain variance in pollution estimates and life expectancy.
With the promise of street view imagery, ethical considerations and potential biases warrant scrutiny. Ongoing research must assess its efficacy in real-world urban health applications, identifying promise and limitations. In sum, this thesis contributes methodological approaches to urban characterisation using street-level imagery, offering a reproducible framework for urban health researchers to leverage in their investigations.
Version
Open Access
Date Issued
2023-07
Date Awarded
2024-04
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Flaxman, Seth
Ezzati, Majid
Sponsor
Medical Research Council (Great Britain)
Publisher Department
School of Public Health
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)