Classification of hyperspectral images by fusion of multifeature under kernel mapping

The spectral bands have strong relation with land covers both partically and theoretically. Thus it is possible to extract enough spectral features with the help of more efficient data represent methods to distinguish land covers. More pertinent spectral matching methods can be taken to improve the similarity and dissimilarity metric and to improve the performance of the classifiers. A couple of classic and efficient spectral measures, such as spectral angle mapper, spectral correlation mapper, mahalanobis distance, spectral similarity value and spectral information divergence, were selected. Then the RBF Guassian function was used and the spectral measure under the kernel mapping were obtained. A new method based on the fusion of multifeatures under kernel mapping was taken to dig the features of hyperspectral remote sensing data. Profile the similarity between different classes and thus a new classification algorithm was proposed. At last, this method was applied to a hyperspectral remotely sensing AVIRIS dataset named 92AV3C using the LIBSVM toolbox of MATLAB. The results show that the classification method of hyperspectral images by fusion of multifeatures under kernel mapping can significantly improve the accuracy of the classification. Experimental comparison shows the proposed algorithm can provide better performance for the pixel classification of hyperspectral image than many other wellknown techniques.