Bhattacharyya Distance based Kernel Method for Hyperspectral Data Multi-Class Classification

Based on the framework of support vector machines (SVM) using one against one (OAO) strategy, a new kernel method based on Bhattacharyya distance is proposed to raise the classification accuracy by combining the characteristics of hyperspectral data. The proposed method takes advantage of the non-uniform information distribution of hyperspectral data and makes the band with greater separability play a more important role during the process of classification. Meanwhile in consideration of the intrinsic binary property of each OAO-SVM classifier, we use the Bhattacharyya distance of the corresponding two species as the spectrally weighted coefficients, which ensure each classifier get its own weights of separability and then lower its classification error. In typical AVIRIS data multi-class classification experiments, using the radial basis function as the basic kernel function, the average accuracies of the proposed method are efficiently improved comparing with standard SVM.