Using AI models to detect sinkhole trouble

Researchers at the University of Florida are developing artificial intelligence models to pinpoint early signs of sinkholes before they appear. "I'm always looking for real-world problems," said Minhee Kim, Ph.D., an assistant professor with the Department of Industrial and Systems Engineering, known as ISE. "We're engineers, so we are always thinking, 'What kind of new problem can we solve using data?' And one of the more promising and relatively unexplored research areas is geomatics."

Considered the bridge between earth science and engineering, geomatics refers to the application of geospatial technologies—satellite imagery, GPS and LiDAR (which measures distances to create 3D maps)—to collect and analyze geological data. But Kim and her team seek to expand those models to glean additional data to identify areas that are at risk of sinkholes.

The stakes are high.

Florida is prone to sinkholes because of the soluble bedrock—mostly limestone—under its surface. Rainwater dissolves it over time, creating underground cavities that, coupled with drought and other harsh conditions, cause the land to collapse.

Though numbers vary depending on the report, there are thousands of sinkholes reported annually in Florida by state officials and insurance companies. Sinkhole damage costs the state hundreds of millions of dollars each year, with average insurance claims exceeding $140,000 in 2020, according to the Insurance Information Institute.

Nationally, sinkhole damage over the last 15 years costs at least $300 million per year, according to the U.S. Geological Survey.

Kim partnered with Chunli Dai, Ph.D., an assistant professor of geomatics at UF, to develop an AI smart system to spot early signs of sinkholes by training and analyzing vastly different data from many sources—satellite images, GPS records, soil measurements and weather reports, for example.

With AI, the system is designed to find and match similar environmental conditions so it can adapt and learn about multiple scenarios.

"Our group's main focus is using remote sensing techniques to identify precursory sinkhole signals, as well as collecting geological, hydrological and other environmental information for past sinkholes," Dai said. "The success of this research will be identifying potential sinkhole formation and providing a sinkhole probability estimate."

Ultimately, the research will produce earlier, more accurate warnings about where sinkholes may appear. The data, Kim said, is designed to provide better information for municipal planners and other stakeholders on where to build or reinforce infrastructure.

The deliverable after this three-year project, Kim said, will be an open-source software program for the public that takes geoscience data as inputs and predicts sinkhole risk.

"While we expect the immediate main users of this program to be geoscientists and researchers from other related fields, we hope it becomes a sustainable effort where it is continuously maintained and extended to benefit the general public," Kim said.

And the more they use it, the more the AI model learns.

"We collect groundwater. We collect satellite imagery. We collect many different types of datasets," Kim said of her team, which includes Sanduni Disanayaka Mudiyanselage, Ph.D., and doctoral student Kani Fu. "We are actively collecting geoscience data across Florida, such that this program and our analysis can cover as much area as possible."

The key to this research is data diversity.

"We came up with a new model that specifically targets unique challenges in geoscience datasets," Kim said. "For instance, when weather conditions are poor, and satellite images are covered by clouds, we can incorporate the other data sets and features to infer missing information."

Those inferences may include soil condition data to fill in some blanks—"accounting for differences between soil conditions in Gainesville and Ocala to help estimate sinkhole risk when satellite images are unavailable, for instance."

"Hopefully," Kim said, "this effort won't end after three years, but this model keeps evolving as people contribute more data and more domain expertise."

Provided by University of Florida