Variant identification of Gastrodia elata Bl. using deep learning assisted FTIR

Gastrodia elata Blume's (TM) variant diversity is one of the important factors that produce differences in chemical composition. Due to differences in active ingredients, there are significant differences in the therapeutic effects of different variants of TM. Rapid and precise identification of various TM variants without causing damage is essential for maintaining their quality and efficacy. This study, FTIR spectral data of three TM samples with different variants were collected by FTIR spectroscopy, and the spectral data were classified into FTIR full spectrum (4000–400 cm−1) and FTIR fingerprint region (1800–900 cm−1). The spectral data were analyzed by PLS-DA, OPLS-DA, SVM, BPNN, GBM, and ResNet to build the qualitative models of TM variants. The results show that the deep learning model (ResNet) is better at recognizing and classifying TM variants than traditional machine learning models (PLS-DA, OPLS-DA, SVM, BPNN, and GBM). The ResNet model, using MSC or SNV spectral preprocessing, achieves 100 % accuracy on the training set, test set, and external validation set. This research develops a digital technique for rapid, non-invasive, and precise identification of TM variants.