Nonlinear correlation coefficient based kernel method for hyperspectral data classification

Under the framework of support vector machines using one against one strategy, a novel kernel method based on nonlinear correlation coefficient is proposed to raise the classification accuracy under the most conditions of no ground truth reference map. This method takes into account the non-uniform information distribution of remote sensing hyperspectral data, and assigns nonlinear correlation coefficients as weights for the corresponding bands to make the band with greater correlation information play a more important role during the process of classification. Meanwhile a new estimated reference map based on nonlinear correlation coefficient is proposed to solve the realistic problem that the real one is usually unavailable. The experimental results show that for the support vector machines based on radial basis function, after adopting the proposed kernels, the average accuracy and the overall accuracy in multi-classification are increased by 2.90% and 3.11% with typical parameter configuration and no ground truth reference map, besides the computational time increment is unobvious.