Kernel method based on normalized mutual information for hyperspectral image classification

Support vector machine (SVM) appears to be a robust alternative for pattern recognition with hyperspectral data. However, this kernel-based method does not take into consideration the bio-physical meaning of the spectral signatures. Observation of real-life spectral signatures from the AVIRIS hyperspectral dataset shows that the useful information for classification is not equally distributed across bands. Hence, we propose the spectrally weighted kernel method to assign weights to corresponding bands according to the amount of useful information they contain, and further research shows that using normalized mutual information (NMI) is the better choice for the estimation of the weighted coefficients than mutual information (MI). We perform experiments on image classification of the 92AV3C dataset to assess the performance of proposed method. Results show that the proposed NMI-based spectrally weighted kernels of polynomial and radial basis function outperform the MI-based kernels accompanied with the ground truth map or the estimated reference map.