Iterative Curve Fitting: A Robust Technique to Estimate the Wavelength Position and Depth of Absorption Features From Spectral Data

In this paper, we introduce a robust method coined iterative curve fitting to estimate the wavelength position and depth from spectral absorption features. The technique iteratively fits a curve to a continuum-removed spectrum and subsets the bands based on the minimum of the previous fit until fulfilling a specified threshold for residual error. The minimum of the latest iteration and its substituted reflectance value are then retrieved as the feature wavelength and depth. Two variants of the technique named iterative Gaussian fitting and iterative polynomial fitting (IPF) are presented. The superiority of these algorithms over current methods is demonstrated using four different absorption features between 400 and 2500 nm collected from an array of sandstones in the laboratory. The methods can achieve rmse values of +/- 1.0 nm for the wavelength and 1% for the feature depth. The estimated wavelength position in a hyperspectral sensor with less than 10-nm sampling interval is demonstrated to be in error by at most +/- 3 nm at 95% confidence level. Experiments with varying signal-to-noise ratios (SNRs) indicate the robustness of the technique against noise. The IPF is able to estimate the wavelength of narrow features with an rmse of +/- 2.7 nm at an SNR of 150 : 1 and broad features with an rmse of +/- 4.2 nm at an SNR of 400 : 1. The method, which is embedded in a package named Automated Absorption-based Mineral Spectral Analyzer (AMISA), enables the simultaneous calculation of width, area, and asymmetry of spectral data acquired from imaging and nonimaging sensors.