Classification of groundwater samples in wetlands using selforganizing maps

Groundwater chemistry in groundwater-fed wetlands often is the result of
mixing of discharging groundwater and rainfall in combination with chemical reactions
altering the chemical composition, mostly due to changes in redox conditions. In this study, a
Self-Organizing Map is used to classify chemical groundwater samples of three groundwaterfed
wetlands in Belgium in order to identify the origin of groundwater and to deduce redox
conditions in the wetlands.
The Self-Organizing Map (SOM) algorithm is an unsupervised neural network technique to
represent a multidimensional dataset on a two-dimensional grid in a topology-preserving way,
allowing investigation of non-linear, complex relationships between variables and grouping of
the data (Kohonen, 1995).
The SOM is trained with data from a regional groundwater monitoring network and rainfall
data. The resulting SOM is able to distinguish between samples of different origin or redox
conditions within the regional aquifers. Subsequently, samples of the three wetlands are
shown to the SOM and each sample is classified as having a chemical composition
comparable to rainfall or to one of the regional aquifers.