Identification of Species in the Polygonatum Genus Based on Different Dimensional Spectral Images Combined with Residual Neural Networks in Fourier Transform Infrared Spectroscopy

Abstract

This study proposed the concept of spectral images with different dimensions，which was first applied to the identification of species in the Polygonatum genus，and established a fast and ac⁃ curate identification method. A total of 563 batches of samples from 6 species of Polygonatum were collected. Five preprocessing methods were used based on Fourier transform mid infrared spectrosco⁃ py（FT-MIR）， including first derivative（1st）， second derivative（2nd）， multiplicative scattering correction（MSC）， standard normal variable（SNV）and Savitzky-Golay（SG）. Decision trees（DT）， random forests（RF）and support vector machines（SVM）were constructed. In addition， to avoid complex preprocessing， a deep learning residual convolutional neural network（ResNet）model was constructed to draw spectral images of different dimensions，including 10 datasets of one-dimensional MIR， synchronous，asynchronous， and integrative two-dimensional correlated spectra，three-di⁃ mensional correlated spectra，and three-dimensional correlated spectral projection images，and they were combined with the ResNet model for classification. The results showed that different preprocess⁃ ing methods have different impacts on the model results，and the MSC preprocessing method signifi⁃ cantly improved the accuracy of the DT，RF，and SVM algorithms. The ResNet algorithm based on synchronous two-dimensional correlated spectral dataset had the best modeling effect，with an accura⁃ cy of 100%，small loss value，no need for complex preprocessing，low time cost，and could accu⁃ rately identify species of Polygonatum genus，providing reference for identification in other fields such as food and traditional Chinese medicine.