Infrared spectroscopy combined with machine learning: A fast method for origin tracing and dry matter content prediction of Dendrobium officinale Kimura et Migo

Abstract

It is a key step to evaluate the quality of Dendrobium officinale Kimura et Migo by verifying its geographical origin and quickly analyzing and predicting its component content. To this end, Fourier transform near-infrared (FT-NIR) and attenuated total reflection Fourier transform mid-infrared (ATR-FTIR) spectroscopy were used to characterize the chemical profiles of D. officinale. The two-dimensional correlation spectrum (2DCOS) images improve the spectral resolution, which is particularly important for analyzing the spectral absorption peaks. Moreover, the potential of infrared spectroscopy in tracing and predicting dry matter content (DMC) of D. officinale in different geographical origins was explored. Different spectral data, preprocessing, feature variable selection, and data fusion methods were compared. The partial least squares discriminant analysis (PLS-DA) model based on the original FT-NIR spectrum was used to trace the origin of D. officinale with 100 % accuracy. The FT-NIR data after second derivative (SD) processing combined with long short-term memory (LSTM) regression model could roughly predict DMC (R²_p = 0.8026, RPD = 1.9149, RMSEC/RMSEP = 1.0433), which provides a rich method reference for the study of D. officinale based on infrared spectrum.