A new fusion of mutual information and Otsu multilevel thresholding technique for hyperspectral band selection

Hyperspectral data are a curse with huge dimensionality, high redundancy of spectral information, and are noisy in nature. Hundreds of narrow adjacent bands are present in HS (Hyperspectral) data with high spectral information and it always leads to a computational complexity in space and time. The information theoretic methods are used for hyperspectral band selection to avoid computational complexity. In order to address this issue, the new fusion of Mutual Information (MI) with Otsu (MI_Otsu) threshold method is proposed for hyperspectral band selection by employing three different entropy measures such as joint, conditional, and relative. The proposed approach identifies the probabilities, entropy, and mutual information between two hyperspectral bands. The optimal threshold is obtained using Otsu multithreshold technique and highly informative bands will be selected. In addition, the SVM (Support Vector Machine) classification technique is adapted for further classification of selected bands to analyze the performance of the proposed algorithm. The experimental analysis is carried out using the real-time dataset from the test site 'Indian Pines' in Northwestern Indiana recorded by AVIRIS (Airborne Visible/ Infrared Imaging Spectrometer) sensor that demonstrates the effectiveness of this proposed approach. It is proved that the proposed work shows the competitive performance even with less selected bands and the relative MI_Otsu method shows a higher accuracy of 92.16% with the comparison of joint and conditional MI_Otsu. (C) 2021 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Computers and Artificial Intelligence, Cairo University.