Automatic Pseudo-invariant Feature Extraction for the Relative Radiometric Normalization of Hyperion Hyperspectral Images

Abstract

A new relative radiometric normalization approach is presented based on the spectral profile shape of hyperspectral data. We calculate the spectral similarity value of pixels at the same location using spectral angle mapping. The cumulative moving average and its differential values are used to determine the appropriate number of pseudo-invariant features automatically. Band-by-band linear regression of the pseudo-invariant features is used to refine the radiometric normalization results iteratively. We tested the algorithm using six Hyperion data subset images. The proposed method yielded stable results with similar or better performance than other methods for all test sites, when assessed by visual inspection and quantitative analysis.