We used machine learning trained on years of resident observations to analyze Philadelphia’s streets and identify signs of gentrification.

(The Conversation is an independent, nonprofit source of news, analysis, and commentary from academic experts.)

(Conversation) What does gentrification look like in Philadelphia?

“High-rise modern apartment”

“(A) It has a modern exterior that is completely out of place with the traditional rowhouses that have been here for 100 years.”

“It’s a six- to seven-story high-rise building with an underground garage. Parking costs an extra $200.”

“It has a gray, industrial feel.”

“The houses are just plain ugly. There’s no architectural style at all. Some have maybe two bedrooms, some have one bedroom. And they usually have decks. They’re not for kids or families. There’s a lot of stairs.”

These are some of the expressions that longtime residents of Philadelphia’s gentrifying neighborhoods have used to describe the new construction that keeps popping up around them.

We have Ph.D. candidate in architectural engineering, geography, environmental and urban studies at Drexel University and Temple University in Philadelphia. Working with an interdisciplinary team of professors and students, we recently developed a new way to map gentrification in Philadelphia neighborhoods by combining long-time residents’ accounts, Google Street View imagery, and machine learning.

Identify gentrification using AI

Our team argued that the best source of information about what gentrification looks like comes from the perceptions of people who have lived in gentrifying neighborhoods for many years.

So we held focus groups in three rapidly gentrifying neighborhoods. One in Northeast Philadelphia and two in the River Ward area along the Delaware River north of Center City.

We asked residents to identify visual cues in building design, materials, colors, and landscaping choices that are associated with gentrification.

Many of these residents were able to detail the exact intersections where gentrification-related development occurred over the decades.

We corroborated each location they identified through historical Google Street View imagery. By considering the appearance of these buildings, you can expand the discussion from more general terms like “modern” and “boxy” to things like “presence of bay windows” and “increased floor area ratio,” which refers to how much of the surface area a building occupies on a parcel of land.

When obtaining panoramas of residential building exteriors from Google Street View, we looked at two different time periods: 2009-2013 and 2017-2021.

AI is getting better at spotting visual signs of gentrification. Researchers refer to AI systems that classify landscapes according to certain characteristics, such as appearing “gentrified” or “not gentrified,” as “deep mapping” models.

Deep mapping models use neural network algorithms that can recognize patterns in large datasets. The particular model we used can detect subtle pixel-level differences between two images.

The model has learned approximately how residents distinguish between gentrified and unchanged scenes. We tested the model’s output and found that it was able to separate “gentrified” and “non-gentrified” images with approximately 84% accuracy. This demonstrated that visual cues based on resident observations can be converted into reliable machine learning signals.

Gentrification doesn’t always look the same

As neighborhoods gentrify, more affluent people move in and longtime residents can be displaced by higher rents or loss of housing. Gentrification can also lead to a loss of the “sense of place” that makes neighborhoods feel familiar and like home.

Deep mapping models enable researchers and neighborhood stakeholders to obtain unique data on landscape changes related to gentrification and better understand how gentrification changes the physical environment. With better data, we can use machine learning models to map hotspots of new developments and predict future trajectories of change.

For example, several focus group participants in one neighborhood noted that gentrification is associated with the demolition of older buildings that are more likely to contain hazardous materials such as asbestos and lead. They wondered about the possibility of air pollution. With accurate data about where development is occurring, researchers can model the relationship between new construction and environmental conditions such as air quality.

Additionally, this process can provide legitimacy to neighborhood groups that are aware of changes occurring around them but lack quantitative data to justify their concerns to the media or city government.

By being more clear about how gentrification is defined when classifying images and training machine learning models, researchers can be more transparent about how image data is prepared and prevent personal bias from guiding the models and the patterns they learn.

For example, one study found that gentrification leads to increased greenery. However, some of our focus group participants reported a loss of community gardens and greenery as a result of gentrification. This experience runs counter to common assumptions in gentrification research.

Build trust with training model transparency

Defining how gentrification is perceived by residents allows researchers like us to have greater clarity on how to prepare model data. But even with greater clarity, these AI systems are still essentially “black boxes.” A black box model means that the connections between inputs and outputs are unclear to model users.

One way to make the model more transparent is to apply an additional model called XAI (Explainable Artificial Intelligence). Through XAI, you may be able to better understand which features in your images are more important for model predictions. For example, does the model focus on building windows or the relative height of buildings?

Answering these questions will help researchers and stakeholders trust the model’s predictions.

At the same time, one of us is leading complementary research focused on explaining the reasoning behind the decisions of machine learning models. In Philadelphia and other US cities, street scenes have a dense mix of cars, vegetation, and architecture that can confuse the model. There’s a lot of complex visual information that needs to be analyzed, and it’s a lot of variety. Understanding the internal logic of a model helps ensure that its predictions reflect real neighborhood dynamics rather than extraneous details in the image.

Together, these research directions aim to advance our understanding of how gentrification unfolds on the ground and how AI can help uncover patterns that might otherwise go unnoticed.

Read more articles about Philadelphia and Pennsylvania or sign up for the Philadelphia newsletter on Substack.

This article is republished from The Conversation under a Creative Commons license. Read the original article here: https://theconversation.com/we-analyzed-philly-street-scenes-and-identified-signs-of-gentrification-using-machine-learning-trained-on-longtime-residents-observations-277704.