Just a few years after early radar emerged in the first years of the 20th century, engineers turned the technology toward the ground to detect subsurface objects without digging, laying the foundation for what is now known as ground-penetrating radar (GPR). Over the past century, GPR has evolved into a powerful non-invasive sensing tool used for infrastructure inspection, agriculture, land mine detection, geology, archaeology and countless other specialized applications.
Now, advances in machine learning and artificial intelligence are creating new opportunities to vastly improve the imaging resolution, accuracy and processing speed of GPR. But there is one major obstacle: There are no large-scale, high-quality GPR datasets for AI models to learn from. Collecting and labeling large real-world GPR data is extremely difficult, expensive and time-consuming. Even worse, datasets collected by different GPR systems are often incompatible, making it nearly impossible to interoperate, share and reuse data across the research community.
That’s why Haihan Sun, an assistant professor in the Department of Electrical and Computer Engineering at the University of Wisconsin-Madison, is embarking on a National Science Foundation-funded CAREER Award project to standardize GPR data and create new sensing algorithms to power the next generation of ground-penetrating radar technology.
Sun’s motivation for this work grew out of her postdoctoral research at Nanyang Technological University in Singapore. In her early efforts to explore how machine learning can improve GPR, Sun had to spend long hours in the humid 90-degree heat building outdoor test sites, digging soil, burying targets and collecting data—a hot and strenuous job.
“Every time I was out in the field, I kept thinking: what if we had algorithms that could standardize GPR data across systems?” she says. “That would save researchers an enormous amount of effort and make it possible for the community to build shared datasets, stronger algorithms and better GPR devices.”
Now, Sun is putting that idea into action. Her team will first develop domain-transfer frameworks that convert incompatible GPR data from different systems into a standardized format. That will make it possible to combine data across systems and build large-scale benchmark datasets. They will then use those standardized datasets to create a physics-informed AI framework that can process GPR signals and generate clearer, higher-resolution and more informative subsurface images in real time. This framework will be integrated with GPR hardware so users can “see” underground conditions on site, instead of waiting for slow off-site processing. Altogether, this will allow users to make faster, better decisions in applications where time, accuracy and safety matter.
Sun plans to validate this technology in three real-world applications, including soil moisture mapping, underground crop imaging and aerial drone-based detection of buried explosive ordnance.
Sun will also launch an initiative called OpenGPRxAI—an open-source, community-driven platform for sharing GPR datasets, pretrained AI models, system designs and educational resources developed throughout the project. Her broader vision is to help build a more connected GPR research ecosystem that can accelerate innovation and expand impact well beyond any single lab.
“This project has the potential to transform how human and autonomous systems perceive and interact with the hidden subsurface world,” says Sun. “It will create exciting opportunities far beyond today’s GPR applications.”
Photo by Joel Hallberg