Transfer of Calibrations of Near-Infrared Spectra Using Neural Networks

Abstract

A new approach for multivariate instrument standardization is presented. This approach is based on the use of neural networks (NNs) for modeling spectral differences between two instruments. In contrast to the piecewise direct standardization (PDS) method to which it is compared, the proposed method builds a single transfer model for all spectral windows. The apparently incompatible requirements for a high number of training objects and a low number of standardization samples are addressed by truncating spectra in finite-size windows and assessing a position index to each window. Each spectral window with the corresponding position index constitutes a training object. No prior background correction is required with this method. Both the proposed method and PDS were applied to some real and simulated data sets, and results were evaluated for reconstruction and subsequent calibration. On the studied data sets, the neural network approach was found to perform at least as well as PDS for both reconstruction and calibration.