Feature Map Classifier - a possible approach to morphological/ geological evaluation of terrain
This paper investigates the practicability of new classification approach to image Processing. Landsat TM image of Slovene Coastal area was used to perform lithological classification. Standard classification methods, based on statistical principles do not always give satisfactory results. Therefore a variety of new approaches are being tested in order to achieve better accuraccy. One of the most promising fields is artificial intelligence where artificial neural networks (ANNs) have proven to be usefull. An artificial neural network represents a limited analogy of the neural functioning of the biological brain (Sejnowski, Koch and Churchland, 1988). Several ANN methods have been developed to solve classification problems. The one represented in this article is a combination of two methods: Self Organising Maps and Backpropagation network. This kind of ANN, also called Feature Map Classifier, is not the best in sence of accuraccy but has one big advantage in comparison with other ANNs - it is much more transparent. In comparison with standard approach better results were gained especially in more complicated cases, where classes are not linearly separable. The separability of classes is shown to be one of the most important factors. ANN methods tend to give better results as statistical clustering technique especially in cases when classes overlap and are not easily separable.