While Google Street View (GSV) has been increasingly available for large-scale examinations of urban landscapes, little is known about how to use this promising data source more cautiously and effectively. Using data for Santa Ana, California, as an example, this study provides an empirical assessment of the sensitivity of GSV-based streetscape measures and their variation patterns. The results show that the measurement outcomes can vary substantially with changes in GSV acquisition parameter settings, specifically spacing and direction. The sensitivity is found to be particularly high for some measurement targets, including humans, objects, and sidewalks. Some of these elements, such as buildings and sidewalks, also show highly correlated patterns of variation indicating their covariance in the mosaic of urban space.

Objectives: Despite theoretical interest in how dimensions of the built environment can help explain the location of crime in micro−geographic units, measuring this is difficult. Methods: This study adopts a strategy that first scrapes images from Google Street View every 20 meters in every street segment in the city of Santa Ana, CA, and then uses machine learning to detect features of the environment. We capture eleven different features across four main dimensions, and demonstrate that their relative presence across street segments considerably increases the explanatory power of models of five different Part 1 crimes. Results: The presence of more persons in the environment is associated with higher levels of crime. The auto−oriented measures—vehicles and pavement—were positively associated with crime rates. For the defensible space measures, the presence of walls has a slowing negative relationship with most crime types, whereas fences did not. And for our two greenspace measures, although terrain was positively associated with crime rates, vegetation exhibited an inverted−U relationship with two crime types. Conclusions: The results demonstrate the efficacy of this approach for measuring the built environment.

Despite the substantial number of studies on the relationships between crime patterns and built environments, the impacts of street-level built environments on crime patterns have not been definitively determined due to the limitations of obtaining detailed streetscape data and conventional analysis models. To fill these gaps, this study focuses on the non-linear relationships and threshold effects between built environments and local crime patterns at the level of a street segment in the City of Santa Ana, California. Using Google Street View (GSV) and semantic segmentation techniques, we quantify the features of the built environment in GSV images. Then, we examine the non-linear relationships and threshold effects between built environment factors and crime by applying interpretable machine learning (IML) methods. While the machine learning models, especially Deep Neural Network (DNN), outperformed negative binomial regression in predicting future crime events, particularly advantageous was that they allowed us to obtain a deeper understanding of the complex relationship between crime patterns and environmental factors. The results of interpreting the DNN model through IML indicate that most streetscape elements showed non-linear relationships and threshold effects with crime patterns that cannot be easily captured by conventional regression model specifications. The non-linearities and threshold effects revealed in this study can shed light on the factors associated with crime patterns and contribute to policy development for public safety from crime.

While street view imagery has accumulated over the years, its use to date has been largely limited to cross-sectional studies. This study explores ways to utilize historical Google Street View (GSV) images for the investigation of neighborhood change. Using data for Santa Ana, California, an experiment is conducted to assess to what extent deep learning-based semantic segmentation, processing historical images much more efficiently than visual inspection, enables one to capture changes in the built environment. More specifically, semantic segmentation results are compared for (1) 248 sites with construction or demolition of buildings and (2) two sets of the same number of randomly selected control cases without such activity. It is found that the deep learning-based semantic segmentation can detect nearly 75% of the construction or demolition sites examined, while screening out over 60% of the control cases. The results suggest that it is particularly effective in detecting changes in the built environment with historical GSV images in areas with more buildings, less pavement, and larger-scale construction (or demolition) projects. False-positive outcomes, however, can emerge due to the imperfection of the deep learning model and the misalignment of GSV image points over years, showing some methodological challenges to be addressed in future research.