Development of a street-by-street analysis of spatial factors in street crime using publicly available crime data by census block

Abstract

This study investigates street crime patterns in Osaka City, Japan, by analyzing spatial factors and identifying specific spatial attributes associated with crime occurrences at a resolution finer than census tracts. Although Japan is renowned for its low violent crime rates, evolving social conditions underscore the importance of continuous crime research. Publicly available crime data in Japan, typically aggregated by census units rather than individual crime points, pose challenges for granular spatial analysis. To address this, the study introduces a method that focuses on street segments By generating road topologies from street polygons and incorporating 3D building data alongside pedestrian flow derived from mobile data. a graph convolutional network (GCN) was developed to predict crime occurrences across street segments. Using data on theft of a bicycle and theft from a car, the GCN achieved accuracy comparable to traditional machine learning methods based on census tracts, with a goodness-of-fit exceeding 0.8 for estimated crime distributions. The findings highlight the critical role of micro-spatial features, such as street width and pedestrian flow, in shaping crime dynamics and suggest practical implications for crime prevention through environmental design (CPTED) strategies.