Sinkhole risk forecasting in the Lithuania–Latvia Karst region using artificial intelligence

The Lithuania–Latvia transboundary gypsum karst region is highly prone to sinkhole formation, posing a significant geohazard to infrastructure, agriculture, and groundwater resources. Risk assessment is challenged by sparse groundwater monitoring networks and strongly heterogeneous karst hydrogeology. This study develops an end-to-end, remote-sensing–informed and data-driven workflow to reconstruct missing daily groundwater-level (GWL) records and to forecast monthly sinkhole formation risk. Daily GWL gaps were reconstructed using supervised machine-learning models driven by satellite-derived climate and water-storage variables. Reconstructed signals were aggregated to monthly resolution and translated into sinkhole risk classes using a Random Forest classifier. A defensible operational target was applied at each well using an empirical 90th-percentile threshold (≥4 newly formed sinkholes per month). Model training employed fold-scoped preprocessing and class-imbalance controls to ensure robust evaluation. Across seven wells (2003–2024), models combining groundwater level, seasonal encoding and hydroclimatic features outperformed single-domain baselines, achieving an accuracy of ∼0.96, high-risk precision of ∼0.98, and recall of ∼0.85. Explainable analyses highlight multi-week hydroclimatic preconditioning as the dominant driver, with sinkhole clusters occurring within ±30 days of groundwater-level peaks. By integrating forecasted groundwater and hydroclimatic features with remote-sensing inputs, the framework can be implemented as an operational decision-support tool or dashboard to deliver up-to-date sinkhole risk alerts, supporting coordinated cross-border infrastructure protection and groundwater management. • Remote sensing and AI used to fill groundwater level gaps in karst terrains. • Groundwater–seasonality features yield best sinkhole-risk forecasts. • High-risk months detected with high precision and strong recall. • Workflow supports early-warning in data-scarce karst region.