Combining two-trace two-dimensional correlation analysis and convolutional autoencoder-based feature extraction from an entire correlation map to enhance vibrational spectroscopic discrimination of geographical origins of agricultural products.

This study explored convolutional autoencoder (CAE)-based feature extraction from entire two-trace two-dimensional (2T2D) correlation maps as a promising tool to enhance the accuracy of vibrational spectroscopy-based discriminant analysis. Although 2T2D correlation maps constructed using only a pair of spectra were effective to highlight minute spectral differences, there was an excessive number of features (variables). Thus, only slice spectra at a wavenumber chosen from the map were typically used for discriminant analysis. In this case, exclusion of a huge number of remaining 2T2D features that would be complementary and descriptive for a given analysis was a major drawback limiting accuracy. Therefore, CAE was adopted to extract features from entire 2T2D correlation maps while minimizing information loss. For evaluation, near-infrared (NIR) and Raman spectra of chili pepper samples and NIR spectra of perilla seed samples were employed for hetero- and homo-spectral 2T2D correlation analysis, respectively. Then, CAE-extracted features from the 2T2D correlation maps were used to discriminate the geographical origins of samples using support vector machine (SVM). Accuracy improved by employing CAE-extracted variables in both cases compared with those using slice spectra chosen from the 2T2D maps. Moreover, to provide clearer insight into the models, gradient-weighted class activation mapping (Grad-CAM) identifying the variables significantly contributed to the discrimination was employed in parallel.