Multiscale Windowed Denoising and Segmentation of Hyperspectral Images

This paper presents the effects of multiscale windowed denoising of spectral signatures before segmentation of hyperspectral images. In the proposed denoising approach it is intended to exploit both spectral and spatial information of the hyperspectral images by using wavelets and principal component analysis. The windowed structure incorporated for this method exploits spatial information by making use of possibly highly correlated pixels. In addition to the proposed method, the segmented PCA is also investigated and compared in the experimental results with a proper modification. In the segmentation process, the K-means and fuzzy-ART algorithms are used. Especially fuzzy-ART is a fast learning network and can be used in high dimensional and high volume data such as hyperspectral images. In the experiments it has been shown that multiscale windowed principal component denoising has positive effects on the segmentation/clustering level.