Examining innovative unsupervised learning techniques for automated characterization of complex groundwater systems

This research proposes an innovative approach utilizing geophysical well logging data analyzed with multiple machine learning (ML) methods including, self-organizing maps (SOMs), k-means cluster analysis (CA), and most frequent value-assisted cluster analysis (MFV-CA), to automatically identify lithological variation within a complex groundwater system. The MFV method is introduced to enhance cluster center identification. In MFV, an automatedly weighted Euclidean distance is applied in which closer data points are assigned higher weights that emphasize their proximity within clusters. This method proved to be efficient, however, to ensure robustness, an innovative histogram-based selection method is employed for initial cluster center positioning to minimize the risk of choosing an inappropriate starting value. The proposed methodology is tested on the Quaternary aquifer system in the Debrecen area, Eastern Hungary which is known for its high lithological heterogeneity. The results were evaluated in both 1D and 2D to reveal the vertical and horizontal distribution of the lithofacies. Accordingly, the histogram-based MFV-CA demonstrated exceptional noise rejection capabilities and efficient recognition of lithological clusters. Moreover, the continuous estimation of hydraulic parameters including hydraulic conductivity and critical flow velocity along the hydrostratigraphical units showed a close agreement with the identified lithological units.