Unsupervised classification of Antarctic satellite imagery using Kohonen's self-organising feature map

This paper describes an investigation into the use of Kohonen's self-organising feature map (SOM) for the classification of remotely sensed imagery of Antarctica. The SOM is an unsupervised neural network which is trained using unlabelled input data. The network consists of a grid of nodes and, after training, each node corresponds to a prototype vector in the input data space. In order to use the trained SOM as an image classifier it is necessary to calibrate the grid of prototype vectors whereby the prototype vectors are clustered and these clusters mapped to physical class labels. The K-means iterative clustering technique is demonstrated as a means of performing this clustering. However this method requires the user to specify the number of clusters to be formed. The U-matrix method is investigated as a way of identifying the number of clusters represented by the grid of prototype vectors.