Evaluation of an autoencoder as a feature extraction tool for near-infrared spectroscopic discriminant analysis

Abstract

The utility of an autoencoder (AE) as a feature extraction tool for near-infrared (NIR) spectroscopy-based discrimination analysis has been explored and the discrimination of the geographic origins of 8 different agricultural products has been performed as the case study. The sample spectral features were broad and insufficient for component distinction due to considerable overlap of individual bands, so AE enabling of extracting the sample-descriptive features in the spectra would help to improve discrimination accuracy. For comparison, four different inputs of AE-extracted features, raw NIR spectra, principal component (PC) scores, and features extracted using locally linear embedding were employed for sample discrimination using support vector machine. The use of AE-extracted feature improved the accuracy in the discrimination of samples in all 8 products. The improvement was more substantial when the sample spectral features were indistinct. It demonstrates that AE is expandable for vibrational spectroscopic discriminant analysis of other samples with complex composition.