A variety of data smoothing techniques exist to address the issue of noise in spectroscopic data. The vast majority, however, require parameter specification by a knowledgeable user, which is typically accomplished by trial and error. In most situations, however, optimal parameters represent a compromise between noise reduction and signal preservation. In this work, we demonstrate a non-parametric regression approach to spectral smoothing using a spatially adaptive penalized least squares (SAPLS) approach. An iterative optimization procedure is employed that permits gradual flexibility in the smooth fit when statistically significant trends based on multiscale statistics assuming white Gaussian noise are detected. With an estimate of the noise level in the spectrum the procedure is fully automatic with a specified confidence level for the statistics. The potential application to the heteroscedastic noise case is also demonstrated. Performance was assessed in simulations using several synthetic spectra by traditional error measures as well as the modality of the resulting fits. For the simulated spectra, a bset case comparison with Savitzky-Golay smoothing method via an exhaustive parameter search was performed while the SAPLA method was assessed for automated application. The applicatiom to several dissimilar experimentally obtained Raman spectra is also presented.
Citation: Applied Spectroscopy
Pub Type: Journals
automated smoothing, noise reduction, non-parametric regression, spectroscopy