AI Reads a Decade of Street Photos and Finds a Buffalo Neighborhood Quietly Rebounding

In the warrens of single-family homes that spread east and south of Buffalo’s old Commodore Perry public housing project, the physical evidence of decades of disinvestment has long been written directly onto the houses themselves: peeling paint, sagging porches, boarded windows, and overgrown lots. A new study published in Frontiers in Sustainable Cities now quantifies, house by house, how that visual record has changed over roughly ten years, and its central finding is striking. Using a multimodal large language model to compare Google Street View images captured in 2014 and 2025, researchers at the University at Buffalo found that properties classified as improved outnumbered those classified as declined by nearly ten to one, an unusually strong asymmetry in a city that has lost population for generations.

The research, conducted by urban planners Li Yin and Robert Mark Silverman, focuses on the blocks surrounding the Perry Homes redevelopment, a public-private-nonprofit partnership that replaced the long-vacant Commodore Perry project with new mixed-income housing. The original 27 brick garden-style apartment buildings, opened in 1939, had deteriorated for decades and sat largely dormant, frustrating earlier attempts to spark renewal in an area where Buffalo’s historically underdeveloped East Side meets the Old First Ward. Funding finally arrived under the federal Rental Assistance Demonstration program, which converted the property from traditional public housing to a project-based Section 8 development. In July 2023, the City of Buffalo approved 260 million dollars for the first phase, which demolished the vacant buildings and delivered approximately 405 affordable townhouse-style units.

What makes the study methodologically significant is not simply the finding of improvement, but the tool used to detect it. Traditional evaluations of redevelopment spillover rely on field surveys, resident questionnaires, or census-tract-level statistics, all of which are costly, slow, and prone to recall bias or aggregation problems. Housing quality, vacancy, and socioeconomic change are usually measured at coarse geographic scales that obscure the fine-grained, street-by-street and parcel-by-parcel variation where redevelopment effects actually play out. Street-level imagery has emerged as a promising alternative, but analyzing thousands of images historically demanded deep learning expertise and region-specific training data that most planning agencies could not muster.

The researchers sidestepped those barriers with a web-based tool that retrieves Street View panoramas through the Google API and passes them to Gemini, a large language model capable of interpreting images directly. The workflow had four stages: defining the study area bounded by Interstate 190, Louisiana Street, a rail line, and Republic Street; collecting and pairing images from the two time periods; human review to filter out unsuitable pairs; and automated visual assessment. When an exact historical match was unavailable, the tool searched nearby capture points and adjusted viewing angles to reconstruct the closest comparable view of each property. After screening, 356 paired images of single-family detached homes remained for analysis.

The evaluation rubric was deliberately concrete. Each image was scored on two weighted categories, building exterior and landscape and surroundings, covering components such as siding integrity, paint quality, roof condition, windows, doors, structural integrity, yard maintenance, fencing, driveway surface, cleanliness, and the upkeep of adjacent properties. Most components were rated on a scale from minus 2 to plus 2, with zero indicating a feature’s absence, and design elements given lower weights because they are less indicative of private investment. The model produced three standardized outputs for each image: a composite score, a building exterior score, and a landscape and surroundings score, all normalized to allow direct comparison across years and categories.

Human oversight ran through the entire process. Two researchers and a graduate student independently reviewed a random subset representing about ten percent of the dataset, achieving more than 90 percent agreement with the model’s assessments. In a parallel exercise, eight undergraduates in a Spring 2026 course analyzed sections of the Old First Ward with manually collected image pairs and two multimodal models of their choosing, generating comparisons that helped identify where image quality, vegetation, vehicles, or ambiguous visual features could distort interpretation. That iterative validation proved essential, since the model occasionally misread transitional states, such as a house under renovation with temporary windows on support boards that Gemini initially classified as boarded up.

The quantitative results point to a decade of broad, if uneven, physical recovery. The mean composite change score was 0.3038 with a median of 0.3000, and scores ranged from minus 1.14 to plus 1.39 across properties. Substantial improvement, defined as a score of 0.66 or above, showed up most often as complete siding replacement or extensive repainting, fresh color schemes, and thorough surface restoration. Substantial decline typically involved accumulating dirt, fading and peeling paint, and missing siding components. Hamburg Street recorded the most substantial improvements, with 24 percent of its properties in that category, while more than 78 percent of Alabama Street properties showed little or no change. Every street contained a mix: clusters of green flags interspersed with at least a few declining properties.

The authors interpret this pattern as evidence that anchor-led investments through public-private-nonprofit partnerships can contribute to localized housing stabilization in shrinking cities, where planners increasingly prioritize smart decline, qualitative growth, and stabilization over growth-oriented strategies that population trends no longer support. The spillover effects around Perry Homes appear mixed rather than uniformly positive, suggesting that catalytic investment can shift a neighborhood’s trajectory without transforming it wholesale. Importantly, the study does not yet connect these physical changes to demographic or economic data, so questions about gentrification, displacement, and whether gains accrue equitably to existing residents remain open for future research linking imagery-derived indicators with census, property value, and tenure data.

For the broader field, the study demonstrates that generative AI can serve as a scalable, low-cost diagnostic instrument for monitoring neighborhood change at a spatial precision that conventional methods rarely achieve. The authors emphasize that the approach should complement, not replace, participatory and value-inclusive planning frameworks, and that open-access data policies are needed so that small property owners and underfunded public agencies can use such tools. As image archives deepen and multimodal models improve, the same before-and-after methodology could be deployed in Detroit, Cleveland, and other legacy cities, giving planners a way to verify whether strategic investments are actually rewriting the visible story of a neighborhood, one façade at a time.

Subject of Research: Using multimodal large language models to assess housing redevelopment spillover effects in a shrinking city

Article Title: Neighborhood transformation in a shrinking city: using multimodal large language models to evaluate housing redevelopment spillover effects in Buffalo, New York

Article References: Yin, L., & Silverman, R. M. (2026). Neighborhood transformation in a shrinking city: using multimodal large language models to evaluate housing redevelopment spillover effects in Buffalo, New York. Frontiers in Sustainable Cities, 8, Article 1925798. https://doi.org/10.3389/frsc.2026.1925798

Image Credits: AI Generated

DOI: 10.3389/frsc.2026.1925798

Keywords: shrinking cities, Buffalo, Perry Homes, public housing redevelopment, multimodal large language models, Google Street View, neighborhood change, spillover effects, urban planning, generative AI, public-private-nonprofit partnerships, Rust Belt