Application of Self Organizing Maps to multi-resolution and multi-spectral remote sensed images

In this paper we investigate the performance of the Kohonen's self organizing map (SOM) as a strategy for the analysis of multi-spectral and multi-resolution remote sensed images. The paper faces the problem of data fusion, by extracting and combining multi-spectral and textural features. Moreover we address the problem of low-quantity and low-quality of labelled pixels in the training set, investigating a two-step strategy: in the first step (unsupervised training) we use a large unlabelled data set to train a SOM, in the second step a limited number of labelled data is used to assign each SOM node to one informative class. Self Organized Maps are shown to be effective way to discover the intrinsic structure of data. When the SOM is used as a classifier, as here, a majority voting technique is usually used to associate nodes with informative classes. This procedure allows to obtain a SOM output grid which contains labelled and unlabelled nodes. Particularly in the framework of remote sensing, the unlabelled nodes may be important, since they are associated with new classes present in the image, or with the so-called mixed pixels, which represent an area on the ground composed of more then one land-cover class. Comparing the results of the proposed SOM-based strategy and the results of a supervised network such as SVM we show that the unlabelled nodes of the SOM are associated with high percentage to mixed pixels.