Nonlinear spectral similarity measure

Abstract

A novel method for spectral similarity measure, which is called nonlinear spectral similarity measure, is presented in This work. In this method, all original spectral vectors are, firstly, nonlinearly transformed into a feature space. Next, kernel PCA is used to construct a set of orthogonal coordinate base in feature space. All transformed spectral vectors are projected onto the orthogonal coordinate space. In kernel principal component analysis (KPCA), the nonlinear translation function is implicatively implemented by kernel function. Moreover, all projected spectral vectors are constrained by spectral continuum removal curve. Because of continuum removal curve, various bands contribute the similar measurement differently. The more absorption, the more the contribution in similarity measurement. At last, linear or general linear similarity measure, for example spectral angle mapper, was used to measure the similarity between two nonlinearly transformed spectra. Our experiments show that this method is effective in spectral similarity measure.