Fuzzy C-mean for unsupervised spectral-spatial SVM classification of hyperspectral images

a new unsupervised spectral-spatial classification approach for hyperspectral images is proposed. The approach combines the results of various unsupervised spectral and spatial classification using majority voting to find the set of labeled samples and over-samples the obtained set to increase the number of learning examples. Fuzzy C-mean (FCM) is used for unsupervised classification and interpolation methods are applied for oversampling. The proposed method used the supervised classifier support vector machine (SVM) with multi-feature kernels to integrate the spatial and the spectral information. In this work, we presented each pixel by a spectral vector illustrating all the spectral information and two spatial vectors presenting textural features and attributes computed by Extended Multi-Attribute Profile (EMAP). Experiments are conducted on AVIRIS "Indian Pines" data set. It was found that the proposed method provided more accurate classification results than FCM and classification without oversampling.