New feature selection method for EO-1/Hyperion image classification: a case study of Subei region, China

Hyperspectral remote sensing can provide tens, even hundreds of spectral bands imagery, which helps us detect the diagnostical spectral characteristics of detected objects. However, there is relatively high correlation between different bands and much redundancy in hyperspectral data sets. Therefore, one of the most important procedures before application is to select optimal bands for extracting information from hyperspectral data effectively. In this paper, we first introduce the characteristics of EO-1/Hyperion, and apply several important pre-processing procedures to Hyperion L1R data, such as radiometric calibration, destriping, smile correction etc. Then we apply spectrum reconstruction approach to feature selection, which uses several basis functions and corresponding spectral intervals to describe the spectrum extracted from Hyperion hyperspectral data sets in Subei region, China. The feature selection method based on spectrum reconstruction is incrementally adding bands to the initial bands, followed by adjustment of band widths and locations. At last, we aggregate several Hyperion bands into a new simulated band in each interval and apply Maximum Likelihood Classification (MLC) method to it. The overall accuracy of classification is 92% compared with in situ measurement, which supports the validity of this feature selection method.