| Abstract: | The limited availability of labeled training samples for supervised hyperspectral image classification decrease the accuracy of supervised classifier. To overcome this difficulty, this paper introduces a new supervised classification approach for remotely sensed hyperspectral image data which process accurately with limited number of training samples and integrates the spectral and spatial information via composite kernel. In fact, the developed method introduces, in the learning step, new examples oversampled from the available limited training set by using interpolation techniques. First, each pixel will be presented by two vectors: spectral vector containing all the spectral information and spatial vector including contextual information extracted using Extended Multi-attribute profiles (EMAP). Then, an interpolation technique is used to generate new training samples, using the limited available data. Finally, a support vector machines (SVMs) with composite kernel is efficiently trained to generate the classification map. The proposed classification approach is experimentally evaluated using both simulated and real hyperspectral data sets, exhibiting higher performance when compared with the classification without oversampling. The integration of interpolation methods with SVMs, combined with the use of spectral and contextual information, represents an innovative contribution in the literature. This approach is shown to provide accurate classification of hyperspectral imagery with limited number of training samples.
|