Band selection and its impact on target detection and classification in hyperspectral image analysis

This paper addresses unsupervised band selection for hyperspectral image analysis. The proposed approach is based on high-order moments. Such moments indicate the deviation of probability distribution function of an image from the Gaussian distribution, so the selected bands have higher chances to contain important target information. Since the bands with close moment values can be very similar, a band similarity measurement is incorporated into the band selection technique to further select most distinct bands using the criterion of divergence. The number of bands to be selected is pre-estimated using a Neyman-Pearson detection theory-based eigen-thresholding approach. The performance of such a band selection technique is evaluated by the detection and classification performance using the selected bands, i.e., the capability of preserving the target information in the original image data.