| Title |
Sinkhole risk forecasting in the Lithuania–Latvia Karst region using artificial intelligence |
| Authors |
Samalavičius, Vytautas ; Bikše, Jānis ; Zaslavsky, Ilya ; Lekstutytė, Ieva ; Arustienė, Jurga ; Žaržojus, Gintaras ; Kunsakova, Assemzhan ; Retike, Inga ; Gadeikienė, Sonata ; Gadeikis, Saulius |
| DOI |
10.1016/j.ejrh.2026.103372 |
| Full Text |
|
| Is Part of |
Journal of Hydrology: Regional Studies.. Amsterdam : Elsevier B.V.. 2026, vol. 65, art. no. 103372, p. [1-24].. eISSN 2214-5818 |
| Keywords [eng] |
Karst aquifers ; machine learning ; sinkholes ; remote sensing ; geohazard ; time-series |
| Abstract [eng] |
Study region 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. Study focus 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. New hydrological insights for the region 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. |
| Published |
Amsterdam : Elsevier B.V |
| Type |
Journal article |
| Language |
English |
| Publication date |
2026 |
| CC license |
|