A completely fuzzy classification chain for multispectral remote sensing images

In this work a new classification algorithm that uses FNP mixed with a pyramidal approach is proposed. The prototypes of each class are generated by means of FCM with a FNP initialization. The aim of the work is to improve the performances of the usual non parametric classifiers by extracting the maximum information from the training pixels and from the pixels to be classified. This is done by using both the high spatial-correlation between pixels and the confidence levels, given by the fuzzy algorithm. Results are presented that show the improvement obtained by applying the proposed method to multispectral image classification.