Multiscale local covariance based feature extraction for segmantation of hyperspectral images

In this work, multiscale local covariance matrices are proposed in the feature extraction step of unsupervised segmentation of the hyperspectral images. Producing groundtruth information for hyperspectral images is very expensive and time consuming process. For this reason, segmentation without label information brings an important advantage for easier analysis of the hyperspectral images. Proposed approach integrates the multiscale principal component analysis and modified local covariance matrices methods in feature extraction phase. To take advantage of employing both spatial and spectral information together, sub-cubes are extracted with a windowed structure for each pixel in the hyperspectral scene. Positive effects of the proposed approach on the segmentation accuracies are proven with the comparative experiments.