Improvement of self-organizing maps with growing capability for goodness evaluation of multispectral training patterns

In this paper, self-organizing maps (SOM) with growing capability are proposed to evaluate the goodness of multispectral training areas selection that would be used in supervised classification processes. The SOM model used in this study is the Growing Cell Structures (GCS) neural network. Some modifications of the original GCS training algorithm are proposed in order to make easy the physical interpretation of their parameters. In addition, several visualization methods have been implemented with the aim of displaying the trained GCS networks. The performances of the modified GCS model have been investigated through a large number of experiments. They have been carried out using multispectral data registered by ETM+ sensor (Landsat 7), to discriminate land cover categories. The results confirm the excellent behavior of the GCS modified training algorithm to evaluate the quality of the selected training patterns, their viability for feeding supervised classification models and their refining.