Date Published: March 13, 2019
Publisher: Public Library of Science
Author(s): Lazar Ilic, M. Sawada, Amaury Zarzelli, Haroldo V. Ribeiro.
Gentrification is multidimensional and complex, but there is general agreement that visible changes to neighbourhoods are a clear manifestation of the process. Recent advances in computer vision and deep learning provide a unique opportunity to support automated mapping or ‘deep mapping’ of perceptual environmental attributes. We present a Siamese convolutional neural network (SCNN) that automatically detects gentrification-like visual changes in temporal sequences of Google Street View (GSV) images. Our SCNN achieves 95.6% test accuracy and is subsequently applied to GSV sequences at 86110 individual properties over a 9-year period in Ottawa, Canada. We use Kernel Density Estimation (KDE) to produce maps that illustrate where the spatial concentration of visual property improvements was highest within the study area at different times from 2007–2016. We find strong concordance between the mapped SCNN results and the spatial distribution of building permits in the City of Ottawa from 2011 to 2016. Our mapped results confirm those urban areas that are known to be undergoing gentrification as well as revealing areas undergoing gentrification that were previously unknown. Our approach differs from previous works because we examine the atomic unit of gentrification, namely, the individual property, for visual property improvements over time and we rely on KDE to describe regions of high spatial intensity that are indicative of gentrification processes.
Working in London, UK, Glass (1964), coined the term ‘gentrification’ to describe the process whereby the working class is displaced by upper classes (the gentry) in urban space . Class displacement resulting in gentrification has been observed in many western cities [2–4]. However, the conceptualization of gentrification has expanded over time to focus on the driving factors of the process such as culture and consumption [5–7], political economy and production [8–11] and, indeed, the definition has extended to include different groups of gentrifiers ranging from marginal  and middle class  to the working class . Current thought on gentrification is intersectional, examining how various discourses are implicated, such as bodies  or industrial spaces  and, as such, gentrification has been identified as taking place where it was once not seen as possible . Recent years have witnessed an expansion of the topic into tourism , planning and policy impacts , environment [19,20], and many other specialized topics [21–26]. The Economist tweeted a quote by Dyckhoff claiming that gentrification is “the most significant force in Western cities in the second half of the 20th century.” [27,28]
Herein, we adopt a deep mapping approach and employ a Siamese-CNN (SCNN) to detect improvements in the frontage quality (building structure plus the front of the property) of individual properties using imagery through time from GSV in Ottawa, Canada. Our approach differs from previous works because we examine the atomic unit of gentrification, namely, the individual property, for visual improvements over time that are indicative of gentrification processes. Because individual properties can undergo a visual gentrification-like change anywhere in geographic space, the SCNN becomes a detector and we rely on mapping the SCNN’s detections to identify clusters of visual changes in space. Our reasoning is as follows: If there is a spatial concentration (a high intensity) of increasingly positive changes in the physical appearance of several properties in close proximity, then this spatial concentration is indicative of a gentrification-like process. However, any given region of high spatial intensity cannot definitively confirm gentrification as the causal factor: to do so would require information on the socioeconomic and cultural changes within and around that space. Nevertheless, as discussed in the introduction, housing stock re-investment and the visual changes that ensue are an acute indicator of gentrification . Therefore, while not confirmative, regions exhibiting a high intensity of visual property upgrades provide a spatial hypothesis that can be tested by examination of other factors implicated in the gentrification process.
SCNN-FC-8 detected 3483 instances of gentrification-like visual changes at a total of 2922 unique locations. When multiple visual changes were detected for a property, we retained the most recent date when producing KDE maps. The KDE maps of the detected results and the building permits exhibit very similar patterns (Fig 5). Two notable differences in the patterns are labelled in Fig 5. The first difference (labelled x Fig 5A) is due to false positive detections that were identified because of a change in the GSV camera, whereby the images along one street were offset for the same geographic locations between 2007 and 2009. The second large difference (labelled x’ on Fig 5B) represents permits within a multi-hectare redevelopment of the city’s ageing cattle dome/football stadium, the majority of which could not be seen in GSV. A difference map of the standardized intensities highlights both x and x’ as spatially distinct regions of differing intensities between the KDE of visual changes detected by SCNN-FC-8 and the KDE of building permits (Figure F(A) in S1 Appendix).
Despite the pervasiveness of gentrification in modern cities, the focus on social, economic and cultural discourse around the phenomenon has led to a neglect in the development of methods that quantify the temporal and spatial evolution of the phenomenon itself. Typically, to identify gentrification, census data is analyzed over time to identify socioeconomic changes in census tract structure [11,50,70,71]. However, as our maps illustrate, gentrification-like visual processes are often localized, particularly in the initial stages and have no natural respect for artificial census boundaries. In general, census data places restrictions on the ability to quantify the process of gentrification at arbitrary temporal and spatial domains.
By taking a deep mapping approach to detecting gentrification, we have shown that it is possible to indicate precisely where and when gentrification processes are happening in an urban area. By focusing on the detection of gentrification-like visual changes to individual properties, SCNN-FC-8 results could be aggregated to any arbitrary geography and thereby specify the proportion of any arbitrary spatial unit that has gentrified. One advantage of doing so would be to test scale and zoning effects of the MAUP on results that are tied to a set of artificial boundaries (such as census tracts or neighbourhoods). For example, with SCNN-FC-8, mapped results are able show if two blocks gentrify around a boundary. This could aid in validating or decomposing the results of census-based inferences about gentrification in urban areas. While the SCNN-FC-8 model was developed and tested in the context of Ottawa, we believe that similar results can be reproduced in other urban contexts within peer countries. It remains to be seen what or if similar results can be replicated in different contexts and the degree to which we can detect different kinds of gentrification (such as slum gentrification). Definitions and visual indicators of gentrification can differ based on cultural or architectural norms across different countries. With a relatively small but regionally consistent training dataset, other north-American locales would be able to train the SCNN-FC-8 architecture using a transfer learning approach.