Speaker
Description
Karst aquifers are complex systems characterised by nonlinear recharge responses and heterogeneous groundwater flow. This study investigates whether unsupervised machine learning can independently distinguish deep karst from shallow karst (fluviokarst) in the northern Croatian Dinaric region.
A multiparametric dataset comprising 5-year records of spring discharge, water temperature and precipitation, complemented by hydrochemical and stable isotope observations, is used to derive descriptors of groundwater response, storage, interaction between rock and water, as well as residence time. Principal component analysis and unsupervised clustering are applied without prior information on the hydrogeological classification of the monitored springs. The resulting data driven groups are compared with an independently established conceptual hydrogeological model.
The analysis examines whether distinct clusters correspond to springs draining the deep karst aquifers of the Kapela Mountain massif and the downstream shallow karst system. The contribution of hydrodynamic and hydrochemical variables is evaluated to determine which parameters provide the strongest discrimination between aquifer types.
By combining hydrogeological knowledge with machine learning, this study explores the potential of environmental time series and hydrochemical data for objective classification and data driven characterisation of complex karst groundwater systems.