Abstracts Geography

EXAMINING NEIGHBORHOOD HEALTH AND BUILT ENVIRONMENT: A DEEP LEARNING APPROACH WITH REMOTELY SENSED IMAGERY

by Zhuo Chen

The places we live, study, and work have major impacts on our health. Except for some diseases that are caused by genetic defects, most of the diseases can be prevented by avoiding exposure to detrimental environments and maintaining a healthy lifestyle, which is highly determined by the built environment in our neighborhood. Thus, measuring the built environment and identifying its health-related elements become important for public health policy and urban planning. Conventional measurement such as questionnaire surveys, however, is time-consuming and not cost-effective. With the state-of-art deep learning and computer vision technology, combined with imagery-based spatial data, we have new opportunities for examining neighborhood health and built environment effectively at a large scale. This dissertation develops and demonstrates the effectiveness of identifying built environment features using deep convolutional neural networks (DCNN). The DCNN-extracted deep features of satellite images and street views show promising performance in modeling the prevalence of obesity at the census tract level. The important deep features that contribute to obesity modeling are identified from the heatmap in the images, including the pavement of roads, decent residential buildings, and amusement parks. The correlation analysis with conventional datasets also suggests that these DCNN-extracted deep features can capture the environment and socioeconomic factors such as PM2.5, employment, poverty, and income. In addition to the public health sector, the research also contributes to the communities of the general public, policymakers, and researchers, by providing a new effective way of analyzing the built environment.