How well do street view images predict crime rates in London? A comparison with social and macro-level environmental features

In research on the causes of crime, geographic context is considered important in relation to how neighbourhood features influence crime. These features include the social and physical environmental features. Historically, measuring the impact of the physical environment – especially its micro-level characteristics – on crime has been challenging due to the lack of fine-grained quantitative data. Recent advances in computer imagery have enabled researchers to extract structured data from street view imagery, creating new opportunities to quantify features of the physical environment at this scale – particularly those visible from the streetscape perspective. However, the predictive power of these features, and particularly how they compare to more traditional sources of neighbourhood data, remain underexplored. Conducting the analysis across a large urban area also presents a significant challenge. To address these gaps, this study uses a stratified random sampling technique (stratified by classes of socio-economic deprivation/affluence) to select and extract data on micro-level environmental features from Google Street View (GSV) images. These are studied alongside other social and macro-level environmental data for 1000 Lower Super Output Areas (LSOAs) in London, using Random Forest as the core predictive model, with Ordinary Least Squares (OLS) and Geographically Weighted Regression (GWR) serving as supplementary tools to predict and analyse crime rates at an additional 500 randomly sampled LSOAs. While ‘social and macro-level environmental features’ – specifically renter occupancy rate, the number of POIs, and transport accessibility scores – were generally the most important predictors of crime, for certain crimes, such as criminal damage and arson, incorporating micro-level environmental features improved model accuracy. Overall, models incorporating spatial information in the relationships between environmental and social features and crime rates, outperformed other models. This underscores the importance of considering spatial heterogeneity when analysing features influencing crime.