An Unsupervised Band Selection Method Based on Overall Accuracy Prediction

This paper proposes an image classification accuracy prediction based unsupervised band selection method for hyper spectral image classification. The key of this method is the prediction of overall classification accuracy for each spectral band with no ground truth or training samples. Under the hypothesis of Gaussian Mixture Model (GMM), we build the explicit expression between the overall accuracy and the distribution parameters of each class, which is denoted as the overall accuracy prediction equation (OCPE). Then, by employing the unsupervised mixture models learning algorithm to predict these distribution parameters, the overall accuracy is computable on the basis of the OCPE. Then, the candidate band subset is obtained by selecting the bands with relatively high overall accuracy. Finally, we use the divergence based band decor relation algorithm to further remove the redundant bands. Real hyper spectral images based experiments show that our band selection method is effective in comparison with other three well-known unsupervised band selection techniques.