Classification and segmentation of hyperspectral images with joint usage of spectral and spatial information

Hyperspectral imaging is developing remote sensing technology which allows to use hundreds of narrow and adjacent bands. Hyperspectral sensors operate between visible region and long wave infrared region on electromagnetic spectrum and able to obtain hundreds of bands. While previous technology, multispectral imaging systems contains generally between 4 and 7 bands approximately between 300 and 400 nm wavelength, hyperspectral imaging systems acquire hundreds of bands between 10-20 nm wavelength. Hyperspectral imaging technology is used for different purposes day after day in different scientific disciplines, especially in the geoscience. Hyperspectral imaging is also used in lots of scientific areas ranging from medicine, chemistry, forestry, agriculture to urban planning, target detection etc .. In the literature several supervised and unsupervised classification methods have been applied to hyperspectral images in order to make them more comprehensible. In this work, unsupervised classification methods have been utilized for unlabeled images and supervised classification methods have been used for images which have ground-truth information. Dimension reduction methods have been used to reduce both time and computational complexities. Band selection and feature extraction methods are widely used for dimension reduction process on hyperspectral images. In this work, principle component analysis (PCA) and kernel principle component analyses (KPCA) are used in dimensionality reduction phase. These methods also make advantages for data projection on Eigen-space. Joint usage of both spectral and spatial information represents more convenient approach in classification. In this thesis, two different feature extraction methods proposed by using spectral-spatial information. The first one, local covariance based feature extraction, defines local covariance matrices for each pixel in the hyperspectral scene. The other one uses local covariance matrices incorporating the multi-resolution analysis (MRA) for feature extraction. After feature extraction phase, pixels in the scene can have more discriminative features. K-means (KM), fuzzy C-means (FCM), Gustafson-Kessel (GK) and expectation maximization (EM) clustering methods have been used for unlabeled images in unsupervised classification. Support Vector Machines (SVM) random forest (RF) classification methods have been used for labeled hyperspectral images. Obtained segmentation and classification maps have been evaluated by using objective and statistical criteria comparatively.