Crime Mapping from Satellite Imagery via Deep Learning

Abstract: Ensuring urban safety is an essential part of developing sustainable cities. An urban safety map can assist cities to prevent future crimes. However, mapping is costly in terms of both time and money due to the need for manual data collection. On the other hand, satellite imagery is becoming increasingly abundant and accessible with higher resolution. Given the outstanding success deep learning has achieved in the field of computer vision and pattern recognition over the past 5 years, in this paper we attempt to investigate the use of deep learning to predict crime rate directly from raw satellite imagery. To this end, we have trained a deep Convolutional Neural Network on satellite images obtained from over one million crime-incident reports collected by the Chicago Police Department. The best performing model predicted crime rate from raw satellite imagery with an accuracy of 79%. To test their reusability, we used the learned Chicago models to predict for the cities of Denver and San Francisco city-scale maps indicating crime rate in three different levels. Compared to maps made from years' worth of data collected by the corresponding police departments, our maps have an accuracy of 72% and 70%, respectively.