Neural networks are powerful tools that could explore the basic structure of environmental data. In this work, the most common artificial neural network (ANN) architectures, multi-layer perceptrons (MLPs), radial basis function (RBF) and Kohonen's self-organizing maps (SOM), are applied in order to assess the quality of the water reservoirs used for the domestic and industrial water supply of the city of Athens, Greece. In parallel, ANN models are optimized and their recognition and predictive accuracy is tested. The data set consisted of 89 samples collected from the three Athenian water reservoirs during a period of 6 months (October 2006 to April 2007). Thirteen metals and metalloids, Fe, B, Al, V, Cr, Mn, Ni, Cu, Zn, As, Cd, Ba, Pb, were determined. For the validation of the optimized ANN models, new data from subsequent sampling campaigns (December 2007) were used. The constructed classification models predicted successfully the origin of the new posterior samples and simultaneously revealed the differences in sample compositions that occurred in that period. Critical comparison of the different architectures in site classification and modeling verified the validity and usefulness of ANNs, as a powerful and effective tool for water quality assessment.