The combination of FT-NIRS and chemometrics realizes the traceability of Lanxangia tsao-ko collection points.

In this study, we realized accurate traceability of 12 Lanxangia tsao-ko (LT) collection points based on Fourier transforms near infrared spectroscopy (FT-NIRS). In order to deeply explore the practical information in FT-NIRS, we extracted 45 statistical features of FT-NIRS and further fused FT-NIRS with the statistical features. The performance of Decision Tree (DT), K-Nearest Neighbor (KNN), and Naive Bayes (NB), three classical machine learning, was significantly improved under the fusion strategy. Suggests that statistical features can be used as characterizing variables to improve modelability. The multimodal model 1D-2D-GRU-CNN-Attentio (1-2-GCA) constructed based on Gated Recurrent Unit (GRU) and Convolutional Neural Networks (CNN) incorporates the fusion of features from 1D dimensional sequences and 2D images, and the LT acquisition point recognition accuracy can all be maintained above 90%. It indicates that the fusion of multiple dimensional features can effectively reflect the sample information and is a potential data mining tool. In addition, the Residual Neural Network (ResNet) model based on two dimensional correlation spectroscopy (2DCOS) and three dimensional correlation spectroscopy (3DCOS) has a strong generalization ability and robustness. It maintains 100% accuracy in many different classification tasks and has the best performance among the five models. In conclusion, this study not only realizes the accurate traceability of LT collection points and provides a non-destructive, fast, and reliable method for the authenticity of LT sources; but the extraction of statistical features and the fusion of multidimensional data also provide a theoretical basis for the in-depth mining and maximization of spectral data.