Band Selection from Hyperspectral Data for Conifer Species Identification

Abstract

Hyperspectral data compression and dimension reducing are very important to computer processing and data transmission. A small number of bands, containing relatively large amount of spectral information, are usually sufficient to many application purposes. Therefore, how to select a small number of bands without loss of much information from all the bands is a critical issue. In this paper, a method of band selection using band prioritization with peak values of sum of 30-eigenvector pertinent to principal component analysis (PCA) was developed. An error back-propagation neural network (NN) algorithm was applied to evaluate the effectiveness of the band selection method in forest species recognition. The results show that, when entering NN with 6–20 bands selected from a total of 161 bands of hyperspectral data for identifying six conifer species, the average recognition accuracy improvement of 11.20% can be obtained using the new band selection method over the method of equal-interval band selection.