Using Deep Learning and Satellite Imagery to Quantify the Impact of the Built Environment on Neighborhood Crime Rates

The built environment has been postulated to have an impact on neighborhood crime rates, however, measures of the built environment can be subjective and differ across studies leading to varying observations on its association with crime rates. Here, we illustrate an accurate and straightforward approach to quantify the impact of the built environment on neighborhood crime rates from high-resolution satellite imagery. Using geo-referenced crime reports and satellite images for three United States cities, we demonstrate how image features consistently identified using a convolutional neural network can explain up to 82% of the variation in neighborhood crime rates. Our results suggest the built environment is a strong predictor of crime rates, and this can lead to structural interventions shown to reduce crime incidence in urban settings. Main Text: Aspects of the built environment such as, building design, street layouts, land use, and environmental disrepair and desolation, have been associated with crime incidence. Differences in the density of these features have been linked to variations in crime rates across diverse geographical settings. However, different built environment characteristics influence particular types of crimes and may work through different mechanisms for crime inducement. For example, the presence of high schools, public parks, vacant lots or buildings can invite gangrelated crimes (1, 2). Also, urban features such as, public transit stations, convenience stores, bars and taverns, and high-rise buildings could increase neighborhood crime rates by attracting transient individuals, facilitating crime through alcohol and drugs, and concentrating disadvantage (2–5). Neighborhood crime has also been associated with structural aspects of the environment linked to the degree of accessibility, and the ease of entry and exit (6). Specifically, major transportation arteries and highways, intersections, alleys and mid-block connections that invite more traffic or enable easy escape have been associated with increased crime and risk of crime in both residential and commercial areas (3, 7–9). Associations have also been noted between accessibility and higher crime when comparing blocks or street segments in high-crime and low-crime neighborhoods (10). In contrast, structural changes in urban neighborhoods have been associated with a reduction in crime rates. For example, a study conducted in London observed that improving lighting in urban streets led to decreases in crime and increases in pedestrian street use after dark (11). In another community intervention in Sarasota, Florida, improvements in city lighting, landscaping, the addition of balconies or porches and residential units to commercial areas combined with new police initiatives for drug dealing and prostitution led to decreases in personal and property crime (12). Additional factors such as street configurations that reduce permeability of cars and cleaning and greening vacant properties were also associated with lower crime rates (13). Although visually identifiable, quantifying the density of these environmental attributes across different geographic regions, populations and over time can be cumbersome. Studies linking neighborhood crime to features of the physical environment have heretofore been conducted using costly and time-consuming onsite visits to count relevant attributes (e.g., the number of liquor stores, vacant lots, and ratings of the level of graffiti or litter in the vicinity of interest) or neighborhood surveys to assess participant perceptions of their neighborhood. The resulting data can therefore be subjective since it relies upon participant or researcher perceptions, and assessment tools that vary across studies. Furthermore, sample sizes for most neighborhood studies tend to be small due to the burden of data collection. The absence of easily accessible data can hamper efforts to identify and quantify the impact of place on crime rates and other relevant public health measures. Here, we demonstrate an accurate, scalable, and straightforward approach that combines a convolutional neural network model and satellite imagery to infer characteristics of the physical environment to assess the degree to which the physical environment can predict variations in crime rates (“predict” here does not indicate forecast of future events). We apply our method to predicting crime rates at the United States census tract level for three cities (Chicago, Illinois; St. Louis, Missouri; and Los Angeles, California) with high crime rates and available geo-referenced crime data. In contrast to existing methods, our approach is low cost, and can produce fine-grained estimates using publicly available data and software.