Wavelet analysis applied to removing non-constant, varying spectroscopic background in multivariate calibration

Multiresolution, the ability to separate signals according to frequency, is one of the main advantages offered by the wavelet transform. However, the coarsening of resolution associated with this method may be problematic in some applications. The 'wavelet prism' (WP) method proposed here can split the signal into different frequency components, which retain the original resolution of the signal. In conjunction with a maximum information gain criterion developed here, this new method can be used to judge and remove the low-frequency non-constant background variation reasonably and automatically. In this paper the theory and background concerning this wavelet baseline correction method are introduced. The method is successfully applied to simulated and real near-infrared (NIR) spectral data to deal with non-constant background for multivariate calibration. Its performance compares favorably with the current signal correction methods for background removal. The new method appears to be an efficient method for removal of non-constant, varying spectroscopic background, leading to a simpler and more parsimonious multivariate linear model. Copyright (C) 2002 John Wiley Sons, Ltd.