Hyperspectral datum classification using kernel method based on mutual information of neighbor bands

Abstract

Listen

Translate article

Under the framework of support vector machines, this paper proposes a new kernel method based on neighbor bands mutual information for hyperspectral datum classification. This algorithm assigns weights to different bands in the kernel function according to the amount of useful information that they contain, which makes the band with more useful information play more important role in the classification. Our research has shown that the band with greater mutual information between neighbor bands contains more useful information, and hence we use the mutual information of each band and its neighbor bands as the weights of the proposed kernel method. The experimental results show that for the support vector machines based on polynomial and radial basis function, after introducing the proposed kernel function, the average accuracy is increased more than 1.2% without using any reference map or increasing much more computational time.