Applying differentiable mutual information to hyperspectral band selection

In this paper, we extend our earlier work by improving a mutual information (MI) based hyperspectral band selection method. Mutual information effectively measures the statistical dependence between two random variables. By modeling ground truth (e.g., a reference map) as one of the two random variables, MI can be used to find the spectral bands that contribute most to image classification. We apply a differentiable rather a histogram-based representation of mutual information to construct the estimated reference map, which results in an automatic solution by gradient searching. Experiments on the AVIRIS 92AV3C data set show that the proposed approach can find the best spectral window, and the bands in this window can be used to construct the reference map satisfactorily.