Segmentation of hyperspectral images using fuzzy approaches

In this paper fuzzy clustering algorithms are utilized for the segmentation of hyperspectral images. For this purpose fuzzy c-means and an extended version of this algorithm, namely the fuzzy Gustafson-Kessel algorithms are used. Because of the high dimensionality in hyperspectral images, the data dimension is reduced using the Discrete Wavelet Transform. The advantage of using fuzzy approaches for the segmentation is that for every pixel fuzzy membership degrees can be obtained. Hereby, a novel method which includes the utilization of spatial information is developed for segmentation with increased accuracy. The method is called `within kernel phase correlation'. Furthermore, it is shown that by two- and three-dimensional Gaussian filtering of the fuzzy membership cube the accuracy can be increased.