In the dense forests and rolling farmlands straddling Lithuania and Latvia, an invisible threat lurks beneath the surface. Gypsum karst—a type of terrain where soluble rock dissolves over time—has created a patchwork of hidden cavities that can suddenly collapse into sinkholes, endangering roads, pipelines, power lines, and even homes. For infrastructure operators, farmers, and water utilities, the unpredictability of these events has long been a costly challenge. But now, a new study led by Vytautas Samalavičius from the Institute of Geosciences at Vilnius University offers a way forward: an artificial intelligence-powered early warning system that could transform how the energy and utilities sectors manage geohazard risk.
Samalavičius and his team developed a remote-sensing-informed, data-driven workflow to reconstruct missing groundwater data and forecast sinkhole formation risk on a monthly basis. Using satellite observations of climate and water storage, they trained machine-learning models to fill gaps in groundwater-level records—a common issue in sparsely monitored regions. The reconstructed data were then used to classify sinkhole risk using a Random Forest classifier, with a clear operational target: predicting when the number of newly formed sinkholes could reach or exceed four per month.
The results are striking. Across seven monitoring wells spanning 2003 to 2024, the models combining groundwater levels, seasonal patterns, and hydroclimatic features achieved an accuracy of about 96%, with high-risk precision at 98% and recall at 85%. What’s more, the analysis revealed that sinkhole clusters tend to occur within about 30 days of groundwater-level peaks, suggesting a multi-week “preconditioning” period where rising water tables weaken underground structures.
“This isn’t just about spotting a hole after it’s already opened,” says Samalavičius. “We’re identifying the hydroclimatic fingerprints that precede collapse—weeks in advance. That window is critical for operators who need to reroute power, reinforce foundations, or divert water flows.”
For the energy sector, the implications are significant. Pipelines, substations, and wind farms in karst-prone regions often face costly shutdowns or emergency repairs when sinkholes appear. A reliable forecasting tool could allow utilities to schedule proactive maintenance, adjust routing plans, or even trigger temporary shutdowns during high-risk periods—saving millions in avoided downtime and repairs.
The study, published in the *Journal of Hydrology: Regional Studies* (Lietuvos ir Latvijos regioninių hidrogeologinių studijų žurnalas), also highlights the potential for cross-border cooperation. Since karst systems don’t respect national borders, a shared alert platform could help Lithuania and Latvia coordinate infrastructure protection and groundwater management—especially as climate change alters precipitation patterns and accelerates karst processes.
Looking ahead, the framework could be adapted to other karst regions worldwide, from Florida to China, where similar hazards threaten critical infrastructure. As machine learning and remote sensing continue to evolve, the integration of real-time satellite data and AI-driven risk models may soon become a standard tool in geohazard management—turning reactive crisis response into predictive resilience.