It is a well-known fact in GIS and remote sensing work that aerial and satellite imagery is most often preferable to manual surveying, which is equal parts slow, tedious, and expensive. It is equally well-known, however, that there are limitations to aerial and satellite imagery - features that are not visible from overhead cannot be usefully detected using typical remote-sensing techniques. The age of the Internet, however, has introduced a new resource that can be used to combat this. Since 2007, the Google corporation has systematically gathered street-level imagery across much of the world and made it freely available for public consumption (Anguelov et al., 2010). In the GIS and remote sensing world, however, this potentially very valuable resource has largely gone untapped, in part due to the need to conduct a manual survey of the imagery, itself a time-consuming and laborious process.
This problem has, however, become eminently solvable given recent advances in computer vision, especially those utilizing artificial neural networks (ANNs). By applying these techniques to Street View imagery, we have developed a tool called TensorView that automatically downloads Street View imagery, processes it using an existing image-classification ANN, and reports which images it believes to meet the criteria specified by the user. We then use the output of this tool as information to set up an appropriate field survey to verify its results. In order to demonstrate the power of this tool, we have chosen a task that is essentially impossible using current remote-sensing techniques - detecting individual abandoned buildings in a small suburban city.
The objectives of this study are thus twofold: first, the development and verification of the TensorView tool and associated workflows, and second, the actual detection of abandoned buildings within the study area.
Overview of Urban Blight
Though there is no simple commonly accepted definition of urban blight, the concept broadly refers to elements of the cityscape that are considered "unacceptable", whether visually or in other ways (Breger, 1967). Most commonly, however, it is restricted in its definition to buildings - specifically, those that are decaying or dilapidated in some way, whether through neglect by the current residents, vandalism, or outright abandonment. In this sense, it is possible to establish a range, as a building can be pristine, a ruin, or at some point in between. At their most obvious, these forces produce the stereotypical abandoned building, typified by boarded-up, broken, or missing windows, graffiti, and (often) structural decay. These buildings can serve as hotbeds for local criminal elements, and also present a significant public-safety hazard due to the risk of fire and structural failure (Spelman, 1993).
Remote sensing of urban blight, however, remains a challenge. To date, only one study of the topic using remote sensing techniques (satellite imagery and multi-criteria evaluation) has been attempted, and the conclusions that were drawn, while appropriate to the 1970s context of decaying cities and suburbanization, are less appropriate given current urban conditions (Tuyahov, Davies, & Holz, 1973). More typically, blight is detected using either manual surveying by government (usually municipal) employees or, in the 21st century, crowdsourced "surveying" by private citizens (Mattioli, 2014; Forrest, 2015). The field-surveying requirement is problematic, as it is recommended that cities maintain up-to-date inventories of blighted buildings in order to better utilize police resources and target negligent landowners.
For the purposes of this study, we use the terms "blighted building" and "abandoned building" interchangeably, and prefer "abandoned building".