Hyperspectral image classification based on non-uniform spatial-spectral kernels

In this paper, several criteria to measure the nonuniform distribution of the spatial-spectral information were investigated in hyperspectral remotely sensed images, then a novel weighted spatial-spectral kernel is introduced. They are proportional to combined (linear / constrained linear and nonlinear) distance measures. Then, the extracted weight vector is applied to both spatial and spectral features. Several spectral-spatial distance criterion including Bhattacharyya distance, the Mutual Information (MI), Spectral Angle Mapper (SAM) and the spatial Markov Random Fields (MRF) energy function are calculated and used in the SVM kernel design. Their combination with three aspects (Linear, constrained linear and nonlinear) regard to convex convergence conditions are examined. The proposed method was implemented for spatial-spectral kernel design. For performance evaluation we used, the data set of Indiana Pines with 190 spectral bands. Comparison of average accuracy, overall accuracy and Kappa coefficient of classification are presented along with class-maps. Experimental results achieved better accuracy and reliability, particularly in terms of limited training samples, in comparison to some classification methods (ECHO and EMP).