||A Study on Public Safety Prediction using Satellite Imagery and Open Data
Data-driven public safety mapping is critical for the sustainable development of cities. Maps visualize patterns and trends about cities that are difficult to spot in data otherwise. For example, a road-safety map made from years’ worth of traffic-accident reports pinpoints roads and highways vulnerable to accidents. Similarly, a crime map highlights where within the city criminal activities abound. Such insights are invaluable to inform sustainable city-planning decision-making and policy. Therefore, public-safety mapping is crucial for urban planning and development worldwide. However, accurate mapping requires longitudinal data collection, which is both highly expensive and labor intensive. Data collection is manual and requires skilled enumerators to conduct. While rich countries are flooded with data, most of poor countries suffer from data poverty. Therefore, cityscale public safety mapping is beyond affordable to low- and middle-income countries. Thus, taking manual data collection out of the equation will quicken the mapping process in general, and make it possible where it is not. Recent advances in imaging and space technology have made high-resolution satellite imagery increasingly abundant, affordable and more accessible. Satellite imagery has a bird’s-eye/aerial viewpoint which makes it a rich medium of visual cues relevant to environmental, social, and economic aspects of urban development. Given the recent breakthroughs made in the field of computer vision and pattern recognition, it is straightforward to attempt predicting public safety directly from satellite imagery. In other words, investigating the use of visual information contained in satellite imagery as a proxy indicator of public safety. In this study, we discuss our approach to public safety prediction directly from raw satellite imagery using tools from modern machine learning and computer vision. Our approach is applied at a city scale thus allowing for the automatic generation of city-scale public safety maps. In this work we focus our attention on two types of public safety maps, namely road safety maps and crime maps. We formalize the problem of public safety mapping as a supervised image classification problem, in which a city-scale satellite map is treated as a set of satellite images each of each is assigned a safety label predicted using a model learned from training samples. To obtain this training data we leverage official police reports collected by police departments and released as open data. The idea is to mine large-scale datasets of official police reports for high-resolution satellite images labeled with safety scores calculated based on number and severity/category of incidents. We validate and test the robustness of the learned models for both road safety and crime rate prediction tasks over four different US cities, namely New York, Chicago, San Francisco, and Denver. We also attempt to investigate the reusability of the learned computational models across different cities. This thesis consists of 5 chapters. Chapter 1 discusses both motivation and background of the study. It also describes how this thesis is organized. Chapter 2 overviews the contributions made in this study which can be summarized as follows: (1) proposing a framework for automatic city-scale public safety prediction from satellite imagery, (2) proposing an automatic approach for obtaining labeled satellite imagery via mining large-scale collections of official police reports released as open data, and (3) introducing five labeled satellite imagery datasets representing four different US cities, and mined from over 2.5 million official police reports. Chapters 3 and 4 describe an extensive empirical study validating the proposed framework. Chapter 3 first introduces a flat image classification architecture that extends an established SVM-based architecture using a novel feature-space local pooling algorithm. This chapter also evaluates the prediction performance of the proposed framework using models learned using the proposed architecture. Chapter 4 continues the empirical study started in the chapter 3 using deep models learned with Convolutional Neural Network-based image classification architecture. The obtained results show that flat models perform modestly compared to deep models which perform reasonably well achieving an average prediction accuracy that reaches up to 79%. This result proves our assumption that visual information contained in satellite imagery has the potential to be used as a proxy indicator of public safety. Finally, chapter 5 summarizes this study and discusses future work directions.
Hokkaido University（北海道大学）. 博士(情報科学)