A fast and effective way for authentication of Dendrobium species: 2DCOS combined with ResNet based on feature bands extracted by spectrum standard deviation.

Dendrobium Sw., as a traditional herb and function food with over 1500 years of history, shows a significant effect in improving immunity and fatigue resistance. However, due of course the large number of species and the quality fluctuating in different species, a fast and effective discrimination method is in need. Recently, spectroscopic techniques combined with chemometrics have become an effective method for low-cost and fast analysis in food and herb. Nevertheless, chemometrics method which based on one-dimensional spectral dataset still encounter the difficulty that can not effectively extract useful information from the spectra. Different from one-dimensional spectra, the two-dimensional correlation spectroscopy (2DCOS) can reveal more detail information of the spectral dataset. Moreover, the appearance of convolutional neural network makes the application of deep learning in image recognition faster and more accurate. In this study, a novel method 2DCOS combined with residual convolutional neural network (ResNet) was used to discriminate the 20 species of Dendrobium. Five feature bands were selected based on spectrum standard deviation (SDD) method in NIR and MIR spectra. Moreover, the models based on full band, total five feature bands, and their fusion-bands had been compared. The results showed that two feature bands 1800-450 cm-1 and 2400-1900 cm-1 displayed 100% accuracy in both training set and test set. And also, the accurate discrimination of 10% external validation showed that these models have good generalization ability. In conclusion, 2DCOS combined with ResNet could be an effective and accurate method for classify different Dendrobium species.