Traditional city planning methods require significant technical expertise and manual work.
A Virginia Tech researcher is working to change that.
New research shows the potential of large language models (LLMs), such as ChatGPT and Google’s Gemini, for assessing the human-made environment using street-view images.
By comparing LLM performance with traditional city planning deep learning methods, the study from the College of Natural Resources and Environment found that LLM-based performance is similar with established approaches. Unlike traditional methods that require technical expertise or manual work, the researchers found LLMs offer a more accessible tool for users, making it easier for policy and planning stakeholders to use these models in small to medium sized cities for managing smart urban infrastructure.
“My goal is to scale down technologies, making them more affordable and effective for smaller cities,” said Junghwan Kim, an assistant professor in the Department of Geography and the director of Smart Cities for Good. “Smart city technologies involve using advanced urban analytics, like AI and data science, to process high-quality data that captures urban environments and how people perceive them. These technologies help us better understand urban issues, such as transportation and health.”
With this new research, it has been shown that generative AI tools to analyze images and detect features like benches, sidewalks, or streetlights automatically.
Previously, researchers had to manually analyze images, which was labor-intensive.
One specific example is evaluating the built environment, like how walkable or bikeable a particular area is. Kim had AI to detect built environment features — benches, sidewalks, trees, and streetlights — all elements that influence how people perceive walkability and safety.
“This democratizes access to advanced tools that were once only usable by experts with coding skills and high-performance computing resources,” Kim said. “However, there are also limitations, such as biases in the AI’s training data, which can cause geographic disparities. For example, these tools tend to perform better in large cities than in smaller towns because of the uneven availability of data for training the AI models.”
The research was published in early October 2024 in The Professional Geographer and was done in collaboration with Kee Moon Jang with the Massachusetts Institute of Technology.
While the tool is very powerful, it can generate hallucinations and make assumptions based on gaps in its training data.
“That’s why it’s important to use these tools carefully, especially in professional settings where accuracy is critical,” Kim said. “I’m still excited about the potential of these tools, not only for my research but also for students and professionals who can now easily access advanced analytics. However, we must remain aware of the limitations and biases that come with using AI in urban planning.”