Improvement of the Classification of Hyperspectral images by Applying a Novel Method for Estimating Reference Reflectance Spectra

Hyperspectral image containing high spectral information has a large number of narrow spectral bands over a continuous spectral range. This allows the identification and recognition of materials and objects based on the comparison of the spectral reflectance of each of them in different wavelengths. Hence, hyperspectral image in the generation of land cover maps can be very efficient. In the hyperspectral classification methods that use a dissimilarity measure for classification, the reference reflectance spectra of each class are usually estimated through averaging the image pixel's reflectance spectra of training data. This estimation method yields a reference reflectance spectrum in which minimize the total sum of squared Euclidean distances between the reference reflectance spectrum itself and the image pixel's reflectance spectra of training data. For this reason, the method is acceptable only for the Minimum Distance algorithm in which is used the squared Euclidean distance for classification. In this paper, we propose a method in which the reference reflectance spectrum is estimated by taking into account the dissimilarity measure that is used in the classification algorithm. Two SAM and JMD classification algorithms have been used to present and implement the proposed method. The evaluation of the accuracy and efficiency of the proposed method has been done by investigating and comparing the results of the classification of SAM and JMD algorithms by considering both averaging and proposed methods. The tests performed on four real hyperspectral images collected by AVIRIS, HYDICE, Hyperion and HyMap sensors show that the proposed method improves classification results, in a manner that the Kappa coefficient of the classification results of four hyperspectral imagery datasets increased by 13.18%, 1.06%, 0.75% and 2.18%, respectively, in the SAM algorithm and 10.79%, 2.17%, 0.34% and 2.4%, respectively, in the JMD algorithm.