The classification of remote sensed images with band extraction and mathematical morphology preprocessing operations

Hyperspectral images contain hundred of images obtained by remote sensing methods and collecting from a large wave length interval. By using these images, for each pixel constituting the image, a continuous spectrum information depending on the wave length is created. The information which is needed for classification is increased by means of these spectrums. This makes possible to augment the classification accuracy at the classification level.As the hyperspectral images have lots of repeating information in itself, on many approaches and applications, the band reduction methods are recently used. Among these methods, the Principal Component Analysis (PCA) is the first step of the preprocessing operation which are mentioned in this thesis.Mathematical morphology operations are used so as to rediscover the spatial information on the bands obtained as a result of PCA. The opening and closing, both are the basic operations of morphology, is the second step by increasing the neighbourhood relation between pixels.In this thesis, the Support Vector Machines (SVM) and Relevance Vector Machines (RVM) are being used as classification methods. Improving of classification accuracy with SVM, mentioned in previous papers, articles and thesis, and comparing SVM and RVM is observed. It is proposed to apply PCA and mathematical morphology operations in order to increase classification performance for both SVM and RVM and also decrease computational load of Relevance Vector Machine (RVM). As preprocessing operations, by using PCA, the number of bands is reduced and by using morphological operations, it becomes possible to use spatial informations of data in additional to the spectral informations that the data has already had. The bands obtained by morphological operations using the results of PCA are processed in RVM. Proposed method shows that the bands obtained after preprocessing is giving better results than the RVM applied to the data directly.