Visualization of multidimensional image data sets using a neural network

The following paper describes the application of self-organizing neural networks on the analysis and visualization of multidimensional data sets. First of all, a mathematical description of cluster analysis, dimensionality reduction and topological ordering is given, taking these methods as problems of discrete optimization. Then, the Kohonen map is introduced, which performs a cluster analysis of the input data and orders its neurons according to topological features of the data sets to be trained with. For this reason, it can also be called a topology preserving feature map. In order to visualize the results obtained during the self-organization process of the network, the standard map has been extended to a three-dimensional cube of neurons, where each neuron represents a discrete entity in the RGB-space. According to the ordering properties of the network, neighbored neurons and thus similar colors refer to data vectors with similar features. This can be shown by training the netwo rk with data sets of characteristic statistical distribution. The application of this technique on multidimensional Landsat-TM remotely sensed image data, namely the analysis of the burning oil fields in Kuwait demonstrates the capabilities of the introduced method. Moreover it can be taken solving general visualization problems of data mapping into a lower dimensional representation.