Measuring Temporally Dynamic Spatial Accessibility using Machine Learned Driving Times

Keywords: spatial accessibility, machine learning, big data
Abstract Type: Paper Abstract

Abstract

Spatial accessibility is used to analyze the distribution of critical resources like food, healthcare, and greenspace across space. While data on population and many of these resources are openly available, real-world driving time data is much more scarce. This means many analyses rely on an assumption of traveling at the speed limit along road networks, but this flawed assumption injects errors into---and potentially biases---our measures of access. In this paper, we explore how machine learning (ML), trained on scarce open driving speed data, can be used to enhance our understanding of access. Using a case study of access to hospital beds in Cincinnati, OH, USA, we seek to answer three key questions: (1) can ML yield more accurate driving time measures for spatial accessibility analyses, (2) how sensitive are existing spatial accessibility measures to the differences between speed limits and predicted driving speeds, and (3) can we leverage ML to gain insights into the temporal dynamics of driving times and access? Our findings demonstrate the power of machine learning to obtain more accurate measures of driving time for spatial analysis, the importance of using accurate and temporally dynamic driving time data when measuring access, and yields practical insights for more accurate spatial accessibility analyses in the absence of such data.