Spectral and spatial classification of hyperspectral images

Recently various developments have been introduced in Remote Sensing Technology. Multispectral sensors have been used for years and provide images with multi bands up to 10-20 bands. The information extracted from the multispectral images are useful and helped in many different applications in the real world, however it may fail in distinguishing between different minerals or sub classes, where hyperspectral images plays an important role. Hyperspectral images are made of hundreds of narrow bands, in most of the cases it can be up to 200 bands. Having this high level of detailed spectral information gives a better distinguishing capability. The conventional classification methods used for multispectral images cannot be applied directly on hyperspectral images due to many reasons. Therefore, many algorithms adjusted or introduced to fit the hyperspectral data. For example, statistical classifiers have difficulties with these data, because calculating statistical parameters for such a huge amount of data is not an easy task. Additionally, the statistical algorithms assume that the data have a specific distribution which contradicts the real-world situation. On the other hand, nonparametric classifies provides good solutions with relatively high and acceptable accuracies. In some cases, these nonparametric algorithms are applied directly on the hyperspectral data or after applying some of the feature extraction methods. K-Nearest Neighbor (KNN) or Support Vector Machines (SVMs) are one of the most widely used and powerful nonparametric methods. Especially SVMs are reported as robust algorithms on hyperspectral data classification even with limited number of training samples. In real world, pixels/samples from neighboring areas are most likely belong to same class. However, classification algorithms exploiting only spectral information cause some noisy like misclassified samples in homogeneous areas. Various approaches have been introduced recently in the literature to improve the classification accuracy and obtain more homogeneous areas in classification maps. One of the powerful approaches is based on integrating the spatial information with the spectral information. In this thesis, we focus on the extended random walker (ERW) algorithm. ERW consists of two main steps; the first step is the spectral classification which is done by any spectral classification algorithm. In this thesis one of the kernel based methods support vector machine (SVM) with the radial basis function (RBF) and the linear kernel function are used in the spectral classification. The second step relies on the results of the classification obtained by SVM and more accurately it depends on the probabilities for each pixel obtained from the SVM algorithm. These probabilities are used to transfer random walker from a segmentation algorithm into a multi class classifier.