Thanks to technologies like Google Street View, anyone with internet access can transport themselves to a sunny sidewalk in Cape Town, the crowded corner outside a Tokyo mall, or a deserted alleyway in London—all from the comfort of their own living room.
But the power of these interactive panoramas goes far beyond sating casual wanderlust: The digitization of city centers is now providing researchers with a treasure trove of visual data on urban change.
According to a study published today in the journal PLOS ONE, a team of geographers has now developed a computer model that analyzes Google Street View images to detect visible changes to the exteriors of individual properties. Though it draws from Google’s incomplete cache of photos, the algorithm achieves a near-human level of accuracy, and has already pinpointed visibly gentrifying areas in Canada’s capital of Ottawa that census-based methods had not yet identified.
“This is a well-executed, solid piece of work,” says César Hidalgo, a computer vision expert at MIT who also researches urban environments through Google Street View images, but was not involved in the study. By monitoring individual properties, he says, the study “advances the field by bringing computer vision methods to understanding urban change at a finer level of granularity.”
Gentrification refers to a phenomenon in which an influx of affluent people reshapes the demographics and aesthetics of an urban neighborhood. The change can be dramatic, and often manifests in the physical renovation of buildings, driving up property values; the higher real estate prices can (but don’t always) displace lower-income, often minority residents. The implications of gentrification remain topics of controversy, but homing in on gentrification as it happens is critical for urban planning and city zoning, and for understanding its impacts on social equity and inclusivity.
However, gentrification has proved difficult to track through traditional methods like census surveys, which are infrequently administered and often draw arbitrary boundaries that don’t match up with actual neighborhoods. As a result, social and political acknowledgement of gentrification can lag behind the process itself, precluding early interventions when they might be most effective.
“Within a couple of years, an entire block can become gentrified,” says study author Michael Sawada, a geomatics researcher at the University of Ottawa. “But we’re not detecting it with census data, and there hasn’t been a means by which to quickly assess a cityscape.”
One way to keep better pace with gentrification, Sawada says, might be to capitalize on its most visible aspects. That’s where Google Street View comes in.
Since 2007, Google has been stockpiling images of cities around the world by dispatching a cavalry of mobile cameras, which have been hooked up to everything from vans to camels, to snap photos at street level. Because the devices make the rounds every few months or years, the database has inadvertently kept tabs on small-scale changes over time—and comprises the most complete archive of street-level imagery on Earth, Sawada says.
A meticulous analysis of Google Street View images could pave the way for watching city’s buildings evolve in real time—but that’s a mind-boggling amount of data for even the most diligent of humans to grapple with. So Sawada and his team instead trained a deep-mapping computer model to analyze 360-degree panoramas of high-density regions of Ottawa taken between 2007 and 2016.
To determine if there had been a “property improvement,” the algorithm examined sequential images of the same location and tallied visible modifications like paint jobs or new fencing. To avoid confusion with garden variety remodeling, the model only flagged gentrification when it noted changes to multiple houses in a concentrated area. At the end of the trials, the model was about as accurate as a human expert sifting through the same images.
When next fed a batch of more than 400,000 images from in and around Ottawa’s Greenbelt neighborhood, the model detected around 3,000 instances of gentrification, the vast majority of which corresponded to places where development and building permits had been granted between 2011 and 2016. In some of these areas, gentrification had already been recognized—but the technology also alerted the researchers to several new and unexpected neighborhoods, Sawada says.
“As much as we may struggle to define gentrification… this shows that you know it when you see it,” says Winifred Curran, an urban geographer at DePaul University who was not involved in the study. “This is a great way to demonstrate that Google and big data can be used to address local concerns.”
Annette Kim, an urban analytics expert at the University of Southern California’s Sol Price School of Public Policy who was not involved in the study, praised the work as an “exciting” new application of Google Street View technology. Kim’s own work, which also utilizes Google Street View images, is geared towards mapping racial and ethnic integration in commercial neighborhoods of Los Angeles. However, she notes that the scope of these techniques remains limited by the data that’s available—and Google Street View is far from comprehensive.
Now in its thirteenth year, Google Street View has uneven global coverage, and entire countries remain undocumented by the tech giant’s street surveyors. Google’s images are also held behind a hefty pay wall, says Hidalgo, the MIT computer vision expert—a fact that could nip some large-scale research efforts in the bud.
“Private sources go where it makes economic sense for them,” Kim says. “If the data we use for public decision making is based on skewed data, we could end up marginalizing people even further.”
Gentrification isn’t limited to the western world, and it doesn’t always look the same. Because the algorithm was trained on Ottawan images, Hidalgo says, that’s where it will perform best—but it could just as easily be stumped by images from another country.
That said, algorithms are malleable, and these types of deep learning techniques remain an important addition to the growing toolkit of urban planning, Sawada says. He and his team are currently tinkering with their model, and hope to adapt the technology to measure other aspects of cities, such as walkability.
“This is a stepping stone,” says Hidalgo, who is also interested in mapping—and ultimately enhancing—urban walkability. “If these techniques are well developed, you could someday have algorithms that don’t just detect visible changes, but also suggest potential improvements to these environments.”
As for tackling the greater issue of gentrification—that’s probably another matter entirely, Curran says. In many cases, by the time Google Street View captures an effect, a property has already changed hands, residents have already been displaced, and a socioeconomic shift is well underway. Other measures will be necessary to truly detect and address gentrification in its earliest stages, she says.
“There’s a lot of stuff that happens behind closed doors before it becomes the thing that you see [on the exterior of a building],” she says. “The visible part is what comes last. So, yes, this is a useful tool, and one we didn’t have before. But at the end of the day, gentrification isn’t about the buildings—it’s about the people.”