Investigating the suitability of self-organizing maps to analyze high-dimensional water quality data sets
Abstract
Water resources management in Europe faces the challenge of the European Water Framework Directive that demands extensive monitoring and evaluating of surface and groundwater quality. As a consequence, there is urgent need to introduce new methods being able to handle multi-dimensional data sets in an effective way. Self-organizing maps proved to be a very powerful tool for this purpose. This type of unsupervised artificial neural networks is now increasingly used as a data-mining tool. They Self-organizing maps project high-dimensional data into a low-dimensional (usually 2-dimensional) space in a non-linear way, aiming at preserving the similarity (euclidian distance) between different data vectors. The resulting map can be processed by Sammon’s mapping in order to display the euclidian distances in the 2-dimensional space proportionally to that of the original high-dimensional data set. This facilitates the interpretation of the map with respect to analyzing patterns in time and space. The self-organizing map was compared to linear methods that are more common in hydrology and geochemistry, i.e., principal component analysis, hierarchical clustering, correspondence analysis, and discriminant analysis, using a high-dimensional water quality data set. The comparison clearly demonstrated the advantages of the non-linear method.
- Publication:
-
EGS - AGU - EUG Joint Assembly
- Pub Date:
- April 2003
- Bibcode:
- 2003EAEJA.....1877L