Self-organizing map improves understanding on the hydrochemical processes in aquifer systems

Abstract

The holistic understanding of hydrochemical features is a crucial task for management and protection of water resources. However, it is challenging for a complex region, where multiple factors can cause hydrochemical changes in studied catchment. We collected 208 groundwater samples from such region in Kumamoto, southern Japan to explicitly characterize these processes by applying machine learning technique. The analyzed groundwater chemistry data like major cations and anions were fed to the self-organizing map (SOM) and the results were compared with classical classification approaches like Stiff diagram, standalone cluster analysis and score plots of principal component analysis (PCA). The SOM with integrated application of clustering divided the data into 11 clusters in this complex region. We confirmed that the results provide much greater details for the associated hydrochemical and contamination processes than the traditional approaches, which show quite good correspondence with the recent high resolution hydrological simulation model and aspects from geochemical modeling. However, the careful application of the SOM is necessary for obtaining accurate results. This study tested different normalization approaches for selecting the best SOM map and found that the topographic error (TE) was more important over the quantization error (QE). For instance, the lower QE obtained from min-max and log normalizations showed problems after clustering the SOM map, since the QE did not confirm the topological preservation. In contrast, the lowest TE obtained from Z-transformation data showed better spatial matching of the clusters with relevant hydrochemical characteristics. The results from this study clearly demonstrated that the SOM is a helpful approach for explicit understanding of the hydrochemical processes on reginal scale that may capably facilitate better groundwater resource management.