ResNet Model Automatically Extracts and Identifies FT-NIR Features for Geographical Traceability of Polygonatum kingianum

Abstract

Medicinal plants have incredibly high economic value, and a practical evaluation of their quality is the key to promoting industry development. The deep learning model based on residual convolutional neural network (ResNet) has the advantage of automatic extraction and the recognition of Fourier transform near-infrared spectroscopy (FT-NIR) features. Models are difficult to understand and interpret because of unknown working mechanisms and decision-making processes. Therefore, in this study, artificial feature extraction methods combine traditional partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM) models to understand and compare deep learning models. The results show that the ResNet model has significant advantages over traditional models in feature extraction and recognition. Secondly, preprocessing has a great impact on the feature extraction and feature extraction, and is beneficial for improving model performance. Competitive adaptive reweighted sampling (CARS) and variable importance in projection (VIP) methods screen out more feature variables after preprocessing, but the number of potential variables (LVs) and successive projections algorithm (SPA) methods obtained is fewer. The SPA method only extracts two variables after preprocessing, causing vital information to be lost. The VIP feature of traditional modelling yields the best results among the four methods. After spectral preprocessing, the recognition rates of the PLS-DA and SVM models are up to 90.16% and 88.52%. For the ResNet model, preprocessing is beneficial for extracting and identifying spectral image features. The ResNet model based on synchronous two-dimensional correlation spectra has a recognition accuracy of 100%. This research is beneficial to the application development of the ResNet model in foods, spices, and medicinal plants.