Over the last several decades urban planners and municipalities have sought to identify and better manage the socioeconomic dynamics associated with rapid development in established neighborhoods. The term "gentrification" has been lingua franca for generations of urbanites who have seen their communities change and property values, and commensurate taxes, shift in ways that can make it difficult for longtime residents to stay. But identifying its unmanaged creep can be a challenge, particularly in densely populated areas, as its visual hallmarks — such as new facades, mixes in building materials and changes in building heights — present differently in different cities and regions.
In hopes of providing a better monitoring system for those seeking to mitigate the negative effects of gentrification, researchers at Drexel University have drawn on the wisdom of community members in Philadelphia neighborhoods that have been affected by it to hone a computer vision program that can reliably identify and track gentrification throughout the city.
Drawing on information from thousands of current and historic images of the city, construction permit records, as well as the input of focus groups from three neighborhoods identified in an analysis of Census data as presenting the socioeconomic shift associated with gentrification, the researchers produced what is believed to be the first "deep mapping" machine learning program that integrates both qualitative and quantitative data to identify gentrification.
The researchers, from Drexel's College of Engineering , recently presented their work and the gentrification identification program they created in the journal PLOS One . They point to Philadelphia's unique and varied architecture and development patterns, housing density and the depth of knowledge from longtime residents as key to training a computer model versatile enough to discern region-specific signs of gentrification.
"While gentrification looks different depending on where it's happening, the people who live in those areas can identify it immediately," said Maya Mueller, a doctoral student in the College of Engineering, who led the research. "Our research is unique in that we ask residents how they identify gentrification in their neighborhood. We then attempted to teach machine learning models to learn from these cues in order to map out where gentrification is occurring."
According to the team, a program like this could help community leaders, urban planners and researchers who are trying to protect residents from being displaced by gentrification, as well as empowering residents who are working to preserve their communities.
"We wanted to open a discussion about how gentrification is changing these neighborhoods," Mueller said. "And through this discussion, develop models that can one day accurately measure the speed and magnitude of these changes."
To create the program, the team connected with residents in three Philadelphia neighborhoods whose socioeconomic shift fit the profile of gentrification and that had been identified through media coverage and the researchers' knowledge of the area. Through a series of focus groups, the team learned about the residents' experiences with gentrification in their neighborhood, the signs of it they recognize in buildings and business corridors and their perceptions about whether and how it has changed access to places and services in the neighborhood.
With this guidance, the team created a list of 16 architectural traits and building qualities indicative of "new-build" gentrification — new construction, as opposed to refurbishing aging buildings — the type most prevalent in gentrifying areas of Philadelphia. The list included things like "boxy" buildings, homogenous design across rowhomes, bump out windows, privacy fences and contrasting mix of building materials with color differences.
"Residents of these areas know gentrification when they see it," said Simi Hoque, PhD , a professor in the College of Engineering, who was a co-author of the research. "In our focus groups they said these buildings 'stick out like a sore thumb.' So, it was then our job to translate the 'sore thumb' into a list of traits that we could use to train our program."
The researchers used the list to label more than 17,000 historic images of Philadelphia neighborhoods from 2009-2013, paired with more recent images of the same places, from 2017-2024, as "gentrification" or "not gentrified."
This information enabled the team to train a neural network machine learning model, called ResNet-50, that learns by comparing subtle variations in training data to identify important characteristics or patterns that it then applies to identify similarities in new inputs.
Through the deep learning processing and the team's manual labeling, the program extracted 1,040 data points that are visual hallmarks of new-build gentrification.
To test the program's gentrification-spotting ability, they showed it new sets of image pairs from around the city. The program was able to correctly identify new-build gentrification in the images with 84% accuracy. To further verify the relative accuracy of the program, the team also compared its audit to permit records for new construction, which has been used as an early indicator of gentrification trends, finding a strong correlation between the two methods.
In addition to producing an accurate program, one of the team's primary goals was to improve transparency and take perception bias out of the process to make it a more reliable tool for urban planners, municipal leaders and community advocates.
"Machine learning models are notoriously 'black box,' so researchers don't fully understand why they produce the predictions they do," Mueller said. "This means machine learning models can learn biases and incorrect ideas and then perpetuate these judgements. It's important that we clearly define how we're training these models, both for ethical reasons and to make these models perform better and more accurately."
According to the researchers, as with any machine learning program, their model would be improved through additional use and exposure to more and varied training data. But it remains a potent tool for researchers seeking to accurately map gentrification trends in areas where reliable permit and development data is lacking.
"With more reliable methods and data on gentrification's effect on the built environment, urban planners can gain insight into how certain types of development result in inequitable effects and organizations can identify neighborhoods that require protection from displacement," Mueller said. "Although we have a way to go, our research team hopes that more concrete measurements on the degree of new development can help to address residents' concerns. Developing this model is one step in the process of producing more useable data."
