A Better AI in the Sky

Through his eyes in the sky, Mehmet Berke Dur '28 gazes down at the world, carefully sifting through patterns of wealth and poverty, looking for communities that might need help.

A computer science major at Trinity University, his summer research project is combining aerial imagery with AI models that he hopes will one day be able to help NGOs effectively distribute aid to people in need. He presented his research on July 23 at Trinity's Summer Undergraduate Research and Internship Symposium. For more stories from the Symposium, go here.

Using a mix of RGB and infrared light, Berke Dur is training a "convolutional neural network" to be able to spot the differences between areas with different incomes.

"I've always been interested in applying artificial intelligence and computing technology in a geographic and geospatial context," Berke Dur says. "This is a mix of computer science, urban studies, and economics, and Trinity is the exact type of interdisciplinary place where I can pursue these things."

Back home in Turkey, Berke Dur watched his uncle's farm in the Aegean region struggle through a harsh drought. As a high schooler, Berke Dur performed his own study on predicting drought rates in the area, determined to use technology to make things better.

Here at Trinity, he's gotten the chance to take the next step in this journey through the Summer Undergraduate Research Fellowship. Fully-funded and supported by an interdisciplinary network of professors, particularly Computer Science Professor Matthew Hibbs, Ph.D., Berke Dur has realized that he's up for the challenges of this type of work.

"At Trinity, you can really connect with professors very closely," Berke Dur says. "But they still give you the independence to make the work 'your thing.'"

This summer, Berke Dur's project has identified a key challenge. Not every area expresses wealth the same way, something he's learned by applying his model to San Antonio, Austin, and Seattle. Something as simple as which neighborhoods are set aside for testing can change how well the model appears to work. "It's a reminder that with AI, how you measure can matter as much as what you build," Berke Dur says.

Next steps for the project involve training the model on bigger cities and a wider variety of areas. One day, Berke Dur hopes to have the capability to apply the technique to smaller towns and rural areas, especially in underdeveloped communities.