What Artificial Intelligence Reveals About Urban Change
Google Street View is like an urban time machine. In the 10 years since it launched, it has captured how neighborhoods have transformed over time—for the better or for worse. What’s not apparent, though, is why some neighborhoods improve and others decline.
To dive into that question, a team of Harvard and MIT economists and computer science researchers is turning to a combination of Street View and artificial intelligence. In a study of neighborhoods’ physical changes and perceived safety, the researchers ran nearly 3,000 images through an algorithm to determine the predictors of neighborhood improvement. While some of the conclusions may not be bombshells for urban experts who study neighborhood change, the researchers say the study, published last week in the journal Proceeding of the National Academies of Sciences, highlights the potential of artificial intelligence to give policymakers and urban scientists a more robust way of testing longstanding theories.
For one thing, the researchers concluded that population density and residents’ education level are two particularly strong predictors of neighborhood improvement, more so than median income levels, housing prices, and rental costs.
The study found that attractive neighborhoods, defined here as appearing safer, are more likely to see improvements. But neighborhoods that appear less safe tend not to fall into further decline, showing mixed support for the theory that when neighborhoods hit a “tipping” point, they will head sharply in one direction.
And finally, the results show support for the spillover effect, the idea that neighborhood transformation is positively linked to its proximity to central business districts and other physically attractive neighborhoods.
Often, these theories are tested using indirect measures of urban change in a small handful of neighborhoods, says Nikhil Naik, a Prize Fellow at Harvard University who led the research and studies the built environment through big data. “Economic successes may be measured by how many new businesses came up,” he tells CityLab. But with the help of machine learning, “we can directly measure the physical change.”
And at a much larger scale. Since 2011, Naik and his colleagues have been asking thousands of people to compare pairs of Street View images from Baltimore, Boston, Detroit, New York City, and Washington, D.C., and assess which one looks safer. Not surprisingly, people ranked images with potholes, broken sidewalks, and dilapidated buildings lower on the perceived safety scale than those with plenty of walkways and green space. Individually, those responses say very little, but his team has fed them into a machine-learning algorithm that can calculate the perceived safety, or “Streetscore,” of any neighborhood street based on its physical attributes.
In this latest study, the researchers ran nearly 3,000 images from those five cities, taken in 2007 and then again in 2014, through the algorithm. Then they calculated the difference in the areas’ Streetscores while accounting for unrelated elements like natural lighting, weather conditions, and the presence of parked vehicles. A positive “Streetchange” score indicates street improvement, while a negative one signals decline. (For accuracy, the scores were checked against human responses garnered from MIT students and participants from a crowdsourcing platform.) The researchers then mapped the Streetchange scores against demographic data from the Census to draw their conclusions.
“What we're trying to do with the tool here is to understand different [aspects] of what makes city better for people, and here it's perceived physical improvement,” says Scott Kominers, a professor at the Harvard Business School and one of the study’s authors. For example, a better understanding of the spillover effect can help urban planners and officials consider how their policies affect not just the immediate neighborhood, but the surrounding communities, as well. “If I build a community center, it may not just improve things for the people who live a block away, but also those in the surrounding rings, so these tools help us understand how big the spillovers are and how far they might move,” Kominers says.
The study is limited in that it mostly looks at cities on the East Coast, which means more research needs to be done to see how applicable the conclusions are to cities around the country—or even overseas. Naik says the next step is to make the data and the tool available to other researchers asking all sorts of different questions. That also calls for improving the algorithm over time as more data is collected and fed into it. Already, they’ve released an interactive map of the five cities showing which neighborhoods and streets show the largest change, positive or negative.
But there’s a caveat. The researchers are careful not to declare causality in their conclusions. They note that neighborhood improvement is positively linked to higher percentage of college-educated residents, but acknowledge that it could be the case that more-educated folks seek out neighborhoods that appear safer.
“The key is that you need the human assessment. This is not a circumstance in which you just set the algorithm and say, ‘Go design a city,’” says Kominers.
“You're designing a city for people, and with people, but the tool makes it possible to work at a much finer resolution and larger scale than you could ever do with just people alone.”
https://www.citylab.com/tech/2017/07/what-ai-has-to-say-about-the-theories-of-urban-change/533211/?utm_source=feed