Using historical data and input from community members, Drexel researchers developed a machine learning program that uses computer vision to identify potentially gentrifying areas.
Over the past few decades, urban planners and local governments have sought to identify and better manage the socio-economic dynamics associated with the rapid development of established areas. The word “gentrification” has been a lingo for generations of city residents who have watched their communities change, property values and corresponding taxes shift, and make it harder for longtime residents to continue living. However, identifying uncontrolled creep can be difficult, especially in densely populated areas, as visual features such as new facades, mixtures of building materials, and changes in building height vary by city or region.
Hoping to provide a better monitoring system for those seeking to reduce the negative effects of gentrification, Drexel University researchers tapped the wisdom of Philadelphia neighborhoods affected by gentrification to hone a computer vision program that can reliably identify and track gentrification across the city.
Utilizing thousands of current and historical city images, information from building permit records, and focus group input from three neighborhoods identified in an analysis of census data as exhibiting socio-economic changes related to gentrification, the researchers created what they believe to be the first “deep mapping” machine learning program to integrate both qualitative and quantitative data to identify gentrification.
Researchers at Drexel University’s School of Engineering recently published their research and the gentrification identification program they created in a journal. pro swan. They point out that Philadelphia’s unique and diverse architectural and development patterns, housing density, and depth of knowledge from longtime residents are keys to training a computer model versatile enough to identify signs of gentrification unique to the area.
“Gentrification looks different depending on where it’s happening, but people who live in the area can easily identify it,” said engineering doctoral student Maya Muller, who led the study. “Our study is unique in that we asked residents how they identify gentrification in their neighborhoods. We then tried to teach a machine learning model to learn from these cues in order to map where gentrification is occurring.”
The team says such programs could not only help local leaders, urban planners and researchers trying to protect residents from displacement due to gentrification, but could also empower residents trying to maintain their communities.
“We wanted to start a conversation about how gentrification is changing these neighborhoods,” Moller said. “And through this discussion, we hope to one day develop models that can accurately measure the rate and magnitude of these changes.”
To create this program, the team connected with residents of three Philadelphia neighborhoods whose socioeconomic changes fit the profile of gentrification and were identified through media coverage and the researchers’ knowledge of the area. Through a series of focus groups, the research team learned about residents’ experiences with gentrification in their neighborhoods, the signs of gentrification they perceived in buildings and business corridors, and their perceptions of whether and how access to neighborhood locations and services had changed.
Based on this guidance, the team created a list of 16 architectural features and building qualities that are indicative of “new construction” gentrification (new construction rather than renovation of aging buildings). This is the most common type in Philadelphia’s gentrifying areas. The list included “boxy” buildings, a homogeneous design throughout the tenement, protruding windows, privacy fences, and contrasting combinations of colored building materials.
“Residents in these areas know gentrification when they see it,” says study co-author Simi Hawk, Ph.D., a professor in the School of Engineering. “In our focus groups, we were told that these buildings ‘stick out like a sore thumb.’ So it was our job to translate that ‘thumb’ into a list of characteristics that could be used to train the program. ”
The researchers used this list to label more than 17,000 historical images of Philadelphia neighborhoods from 2009 to 2013, as well as more recent images of the same locations from 2017 to 2024, as either “gentrifying” or “non-gentrifying.”
This information enabled the team to train a neural network machine learning model called ResNet-50. The model learns by comparing subtle changes in training data, identifies important characteristics and patterns, and applies them to identify similarities in new inputs.
Through deep learning processing and the team’s manual labeling, the program extracted 1,040 data points that are the visual signatures of new luxury homes.
To test the program’s ability to spot gentrification, they showed them a new set of image pairs from around the city. The program was able to accurately identify gentrification of new homes in images with 84% accuracy. To further test the program’s relative accuracy, the team also compared its audits to new construction permit records, which are used as an early indicator of gentrification trends, and found a strong correlation between the two methods.
In addition to creating an accurate program, one of the team’s main goals was to increase transparency and remove cognitive bias from the process, making the program a more trusted tool for city planners, local government leaders, and community advocates.
“Machine learning models are notoriously ‘black boxes’, so researchers don’t fully understand why they produce the predictions they do,” Muller says. “This means that machine learning models can learn biases and erroneous ideas and perpetuate their judgments. For ethical reasons and to make their performance better and more accurate, it is important to clearly define how these models are trained.”
As with any machine learning program, further use and exposure to more and different training data will improve the model, according to the researchers. But it remains a powerful tool for researchers seeking to accurately map gentrification trends in areas where reliable permitting and development data are lacking. “
“More reliable methods and data on the effects of gentrification on the built environment can give city planners insight into how certain types of development have inequitable impacts, and can help organizations identify areas that need protection from eviction. There is still a way to go, but our research team hopes that more specific measurements of the extent of new development will help address residents’ concerns,” Moller said. Developing this model is one step in the process of generating more useful data. ”
In addition to Mueller and Hoque, Isaac Quaye, James Foley, Reeya Shah, and Hamil Pearsall of Temple University, and Xiaojiang Li and Shengao Yi of the University of Pennsylvania contributed to the study. Read the full article here: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0341844
