Methodology for hyperspectral band selection

While hyperspectral data are very rich in information, processing the hyperspectral data poses several challenges regarding computational requirements, information redundancy removal, relevant information identification, and modeling accuracy. In this paper we present a new methodology for combining unsupervised and supervised methods under classification accuracy and computational requirement constraints that is designed to perform hyperspectral band (wavelength range) selection and statistical modeling method selection. The band and method selections are utilized for prediction of continuous ground variables using airborne hyperspectral measurements. The novelty of the proposed work is in combining strengths of unsupervised and supervised band selection methods to build a computationally efficient and accurate band selection system. The unsupervised methods are used to rank hyperspectral bands while the accuracy of the predictions of supervised methods are used to score those rankings. We conducted experiments with seven unsupervised and three supervised methods. The list of unsupervised methods includes information entropy, first and second spectral derivative, spatial contrast, spectral ratio, correlation, and principal component analysis ranking combined with regression, regression tree, and instance-based supervised methods. These methods were applied to a data set that relates ground measurements of soil electrical conductivity with airborne hyperspectral image values. The outcomes of our analysis led to a conclusion that the optimum number of bands in this domain is the top four to eight bands obtained by the entropy unsupervised method followed by the regression tree supervised method evaluation. Although the proposed band selection approach is demonstrated with a data set from the precision agriculture domain, it applies in other hyperspectral application domains.