Explaining Crime Diversity with Google Street View

Abstract

Objectives

Crime diversity is a measure of the variety of criminal offenses in a local environment, similar to ecological diversity. While crime diversity distributions have been explained via neutral models, to date the environmental and social mechanisms behind crime diversity have not been investigated. Building on recent work demonstrating that crime rates can be inferred from street level imagery with neural network computer vision models, in this paper we consider the task of inferring crime diversity through street level imagery.

Methods

We use the Google Vision API, a deep learning image tagging service, to extract objects from sampled Google Street View (GSV) images in each census block of Los Angeles. For each census block we then compute indices for (1) object diversity, (2) diversity related to commonly employed census variables, and (3) crime diversity from reports provided by the Los Angeles Police Department. We then build ordinary least squares and geographically weighted regression models to explain crime diversity as a function of environmental diversity, population diversity, and population size.

Results

We show that crime diversity arises via a combination of environmental diversity (as measured through street view object diversity), household diversity (as measured through the census), and population size. Population size and area of the census block both lend credence to the neutral model proposed by Brantingham for crime diversity. However, environmental and demographic diversity combined play an equally important role in explaining variation in crime diversity.

Conclusions

Our study has two primary implications for research on crime and place. First, Google Street View (via the Google Vision API) can provide important, cost-effective empirical insights to best understand distinct geographic environments of crime. Second, environmental diversity, as measured by image tagging in GSV, was observed to be more predictive of crime diversity (variety of crime types) than commonly used census measures.

Appendix: Handling Synonyms in GSV Object Detection

Here we explore the sensitivity of the diversity index with regards to expanding or collapsing GSV categories. Though some image tags may appear to be synonyms, many are in fact distinct and provide a richer understanding of what is contained within a given image. For example, a family home and apartment are clearly different, so too are row houses that are physically connected as compared to standalone single family homes. While these may indeed be homes or residences that could be combined into a single category, they are unique from one another and may have implications for how offending environments are interpreted. We queried a sample of images with similar tags (or potential synonyms) and reviewed these images in an attempt to decipher differences. For example, it appears yard tends to be the backyard of a residence whereas lawn is the front lawn. Certainly these may be synonymous, but they speak to the view of a given location and how an offender may interpret. Other examples include fence which appear to be metal or commercial fencing and home fencing which appear to be smaller wood fences. They are also distinguished by other characteristics in a given image. A residential home with a fence will be tagged with home fencing whereas a commercial building with a fence will be tagged as fence.

The GSV Vision API attempts to distinguish between these similar but different structures. In fact, the classification is deterministic using a neural network where low level features (image texture, color, edges, etc.) are then mapped to high level features, and finally a prediction. However, certain classes may be similar, for example dwelling and house (though there is something in the image that is causing dwelling to be selected, so the images are somehow distinguishable). To assess the potential sensitivity of the environmental diversity index to combining synonyms into more general categories, we run the following simulation. We simulate 1000 census blocks as follows. In each block we first draw a dirichlet distribution for 20 synthetic categories. We then sample M crimes (where M is uniform from 1 to 1000) according to the dirichlet distribution and allocate them to categories. Next we simulate adding synonym noise by distributing each event to one of 5 random synonyms. We finally compute the diversity index of each of these distributions.

In Fig.  4 we see that even when adding synonym noise, the diversity index is highly correlated (.919) with the original distribution. It is also worth pointing out that both Lentz (2018) and Brantingham (2016) encounter this same issue. In both papers, crime categories themselves are somewhat arbitrary, and different types of clustering could be possible. However they both use the raw categories to compute crime diversity, which is simpler, more reproducible, and less subjective. We prefer to take this approach here, both with the crime diversity index (to match more closely with Lentz (2018), Brantingham (2016)) and the GSV index. We believe this approach is most appropriate given the limited use of GSV Vision API in criminology for purposes of future replication. This approach also avoids the subjective and arbitrary process of assigning image tags to collapsed categories.

Fig. 4

Shannon diversity index of data with simulated synonyms vs Shannon diversity index of collapsed categories