Close

See How Your Peers Are Moving Forward in the Cloud

New research from CDW can help you build on your success and take the next step.

Feb 23 2023
Data Analytics

Baltimore Fine-Tunes Its AI Tool for Spotting Roof Damage

Focusing on building outlines rather than parcels in aerial images could improve the algorithm’s precision in detecting dangerous structures.

The City of Baltimore is refining an artificial intelligence tool that helps prioritize vacant buildings for emergency demolition, specifically working to better recognize roof damage.

Using geographic information systems aerial imagery, the tool looks for dark black pixels by parcel. However, doing so by building outline rather than parcel would reduce the number of shadows mistaken for roof damage.

Data Science for Social Good, a fellowship program at Carnegie Mellon University, trained the AI algorithm last summer. DSSG did so with a data set created in 2018 by eight YouthWorks interns, who flagged roof damage in aerial images of city blocks with vacant buildings. Within its cloud-based, Linux coding environment, DSSG’s model generated a list of 1,000 properties likely to have severe roof problems. Now, Baltimore City Information & Technology is busy ensuring the model operates securely in its Windows environment.

“We’re basically in the process of converting the model into something that adheres to the city’s security standards,” says Henry Waldron, director of research and analytics for the Baltimore City Department of Housing & Community Development. “We’re not able to run it at the push of a button yet.”

Click the banner below to receive customized content by becoming an Insider.

DHCD Sees Room for the AI to Improve

DHCD has greenlit the emergency demolition of about 10 vacant buildings with severe roof damage using DSSG’s model, and officials hope to expand the list of at-risk properties to 2,000 or more and make it public, Waldron says.

While the model produces a risk score for parcels, DHCD also looks at whether a vacant building is free-standing or at the end of a row, which increases the risk of roof collapse due to the lack of structural support.

At the same time, at-risk buildings in the middle of rows are a threat to community development. DHCD will look to fund the stabilization of at-risk vacant buildings amid occupied properties to prevent vacancy from spreading, Waldron says.

DHCD also plans to send the model’s analysis to the Baltimore City Fire Department so its emergency response interface has information about any at-risk buildings when they arrive at a scene, he adds.

Baltimore pays a third party for annual aerial imagery, taken in the winter when trees are bare for greater visibility. The AI model also will be run annually to catch minor developing roof damage that would otherwise become more significant over time, Waldron says.

DISCOVER: The current landscape for AI in state and local government.

DSSG and the City Think the Model Can Be Replicated

Baltimore Chief Data Officer Justin Elszasz connected DSSG with DHCD. Program and city officials visited each other and then began weekly project meetings.

DSSG is interested in refining the tool as well.

“We are exploring some opportunities and are open to other collaborations with agencies interested in expanding or funding this work,” says Rayid Ghani, a professor at Carnegie Mellon and DSSG’s director.

DSSG actually rolled three models into one to develop Baltimore’s AI tool. In addition to the model that counts dark black pixels by parcel in aerial imagery as a proxy for roof damage, a second model factors in data from building inspection notes, permits and vacancy notices. The third model is a neural network trained to review aerial imagery.

“This could be replicable across cities, and other agencies could absolutely use it, especially in cities with aging building or housing stock,” Waldron says.

EXPLORE: The data analytics and sharing tools smart cities need.

ChrisBoswell/Getty Images