Application of neural networks for the classification of ground and surface water observation points using hydrochemical datasets - Identification of geogenic and anthropogenic sources at a former mining site

Valentin Haselbeck1, Jannes Kordilla2, Florian Krause3, Martin Sauter2
1 Abteilung angewandte Geologie, Hydro,-Environmental Geology, Georg-August-Universität Göttingen, K+S AG
2 Georg-Augsut Universität, Abteilung angewandte Geologie
3 K+S AG, Hydro-, Environmental Geology

O 19.3 in Young Hydrogeologists forum

21.03.2018, 13:00-13:15, 3

Growing hydrochemical datasets with many measured parameters and large differences in the measured concentrations creates the demand for a method, that is able to compress information without many losses and subsequently displays it in a clear, simple and comprehensive way. Here, neural networks, a branch of artificial intelligence, is the method of choice. Kohonen's self-organizing maps (SOM) are applied to project the data onto a two-dimensional grid first and then the geometrical relationship of the projected vectors is used to perform a cluster analysis. This two-stage approach is a tool for the hydrogeologist to get a quick and understandable overview over large datasets and present those clearly and comprehensibly.

In this Master Thesis, the SOM is trained, after pre-processing, with the hydrochemical dataset, namely the major ion composition, from a former mining site.  In a second step, the SOM is clustered using hierarchical clustering. At all stages, a sensitivity analysis of the crucial parameters and specifications is performed and the results controlled with quantitative and qualitative measures. Additionally, the final outcome is compared to the outcome of principal component analysis (PCA), displayed in a Piper diagram and used to make those methods "smart". The projection of test data onto the trained map can show its usability for future monitoring of changes in the chemical state. In a last step, the SOM clusters are displayed spatially in the mining area and compared to geophysical measurements and a numerical flow and transport model, modified to meet the purpose of evaluating the outcome of the SOM approach. The clustering result is also discussed and compared to previous studies. All those steps are supplemented with detailed explanations on the functionality of SOM, PCA, clustering methods and hydrogeological modeling.

The SOM clustering approach succeeded in assigning the groundwater samples automatically to a geogenic or anthropogenic source based on their major ion composition. Five different clusters, three geogenic and two anthropogenic were found and verified. 

Piper diagram of the SOM training data colored according to SOM clustering.
Piper diagram of the SOM training data colored according to SOM clustering.



Nguyen, Thuy Thanh, Akira Kawamura, Thanh Ngoc Tong, Naoko Nakagawa, Hideo Amaguchi, and Romeo Gilbuena (2015). “Clustering spatio–seasonal hydrogeochemical data using self-organizing maps for groundwater quality assessment in the Red River Delta, Vietnam.” In: Journal of Hydrology 522, pp. 661 –673.

Vesanto, J. and E. Alhoniemi (2000). “Clustering of the Self-organizing Map.” In: Trans. Neur. Netw. 11.3, pp. 586–600.

Kohonen, Teuvo (2013). “Essentials of the Self-organizing Map.” In: Neural Netw. 37, pp. 52–65.

Choi, Byoung Young, Seong Taek Yun, Kyoung Ho Kim, JiWook Kim, Hyang Mi Kim, and Yong Kwon Koh (2014). “Hydrogeochemical interpretation of South Korean groundwater monitoring data using Self-Organizing Maps.” In: Journal of Geochemical Exploration 137, pp. 73–84.



Export as iCal: Export iCal