Novel non-destructive authentication of nine Dendrobium species using residual convolutional neural network relying on plant images and FT-NIR spectral information

• FT-NIR technology combined with machine learning for non-destructive discrimination. • Two machine learning algorithms: ResNet and SVM were explored to species identification of Dendrobium . • Plant, 1D and synchronous 2DCOS spectral images also used for modeling firstly. • The optimal 32-layer convolutional neural network of the ResNet model based on synchronous 2DCOS images was applied. . Various Dendrobium species used as traditional Chinese medicine have similar appearances but different bioactive component, with significant differences in medicinal and economic values. Many commercially available herbal medicines, including Dendrobium are usually powdered, and commercial fraud by adulterating cheap species in the supply chain often occurs. Therefore, it is necessary to develop accurate and feasible authentication methods for herbal medicines. Currently, non-destructive testing and analysis are gradually becoming a hot issue in various industries. This study attempts to use machine learning techniques for non-destructive combination authentication of different Dendrobium species based on plant photographs (SONY) and Fourier transform near-infrared spectroscopy (FT-NIR). Simultaneously validated the ability of the residual neural network (ResNet) and support vector machine (SVM) models to extract and recognize features from different preprocessed datasets. The results showed that Dendrobium officinale had the highest absorbance followed by Dendrobium thyrsiflorum and Dendrobium crepidatum had the lowest. When the weight decay coefficient λ of the deep learning model based on ResNet is 0.0001 and the learning rate is 0.01, it can identify up to 100% of Dendrobium species. ResNet recognizes feature information of plant images with an accuracy of up to 88.2%. Using flower parts with more recognised features or controlling the consistency of the background may improve recognition accuracy. The dataset of synchronized two-dimensional correlation spectroscopy (2DCOS) does not require preprocessing, and the ResNet model can accurately and quickly extract recognition features. Deep learning models based on ResNet have absolute advantages over traditional SVM models in terms of accuracy and recognition speed. The analytical method proposed in this study may provide new ideas for non-destructive identification of similar species, genuine and fake products, pest and disease characteristics in the field of agriculture.