A THREE-DIMENSIONAL FILTERING METHOD FOR SPECTRAL-SPATIAL HYPERSPECTRAL IMAGE CLASSIFICATION

A suitable filtering preprocessing is beneficial to hyperspectral image (HSI) classification. In this paper, we design a three-dimensional filtering approach for spectral=spatial HSI classification. The associated three-dimensional filter is the coupling of two kinds of kernels. The former is a Gaussian kernel that collects spatial dependency of spectra, whereas the latter is the derivative of Gaussian kernel that reflects the discriminating contribution of spectrum derivative. Through convoluting with such a three-dimensional filtering, it is expected that class-specific spectra of an HSI are more congregated; thus leading to better classification performance in the subsequent classification stage. Experimental results from a real HSI validate that such a three-dimensional filtering is capable of enhancing the classification accuracy of several benchmark HSI classifiers.