Integrating waterborne GPR and group-aware machine learning to resolve lakebed hydrostratigraphy and potential groundwater–lake connectivity for Lake Vadkerti, Hungary

Study Region: Lake Vadkerti is a climate-stressed shallow alkaline lake in the Danube–Tisza Interfluve, Hungary, where lakebed controls on groundwater–lake connectivity remain poorly constrained. Study Focus: We integrated 22 waterborne 200 MHz GPR profiles, 16 upper-lakebed sediment samples, seven interpreted reflectors, processing-aware amplitude analysis and sample-group-aware machine learning to characterize hydrostratigraphic heterogeneity and screen candidate exchange zones. The shallowest consistently traceable sub-lakebed reflector was selected because it provided the closest GPR correspondence to the 20–30 cm sediment samples. Twelve profile-supported sites formed independent validation groups. Five classifiers were evaluated across 48 class-complete leave-three-groups-out combinations. New Hydrological Insights for the Region: Processed amplitude at the shallowest traceable sub-lakebed reflector distinguished Mud from Sand and Transitional sediment. Among five classifiers, tree ensembles performed best. At the 5 m along-profile support, Extra Trees achieved a balanced accuracy of 0.736 ± 0.201; Mud was consistently identified, whereas recall was lower for Sand and Transitional sediment. Incorrect classifications had greater uncertainty than correct classifications. Combining reflector discontinuity, thin-mud support, model support for Sand and proximity to Sand samples identified profile-constrained candidate exchange zones, mainly along the southern and eastern margins and locally along the western margin. These zones are priorities for hydrological testing, not evidence of measured seepage, direction or flux. The workflow is transferable to shallow lakes worldwide where lakebed observations are sparse.