Tanshinone Content Prediction and Geographical Origin Classification of Salvia miltiorrhiza by Combining Hyperspectral Imaging with Chemometrics

Abstract

Hyperspectral imaging (HSI) technology was combined with chemometrics to achieve rapid determination of tanshinone contents in Salvia miltiorrhiza, as well as the rapid identification of its origins. Derivative (D1), second derivative (D2), Savitzky-Golay filtering (SG), multiplicative scatter correction (MSC), and standard normal variate transformation (SNV) were utilized to preprocess original spectrum (ORI). Partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM) models were employed to discriminate 420 Salvia miltiorrhiza samples collected from Shandong, Hebei, Shanxi, Sichuan, and Anhui Provinces. The contents of tanshinone IIA, tanshinone I, cryptotanshinone, and total tanshinones in Salvia miltiorrhiza were predicted by the back-propagation neural network (BPNN), partial least square regression (PLSR), and random forest (RF). Finally, effective wavelengths were selected using the successive projections algorithm (SPA) and variable iterative space shrinkage approach (VISSA). The results indicated that the D1-PLS-DA model performed the best with a classification accuracy of 98.97%. SG-BPNN achieved the best prediction effect for cryptotanshinone (RMSEP = 0.527, RPD = 3.25), ORI-BPNN achieved the best prediction effect for tanshinone IIA (RMSEP = 0.332, RPD = 3.34), MSC-PLSR achieved the best prediction effect for tanshinone I (RMSEP = 0.110, RPD = 4.03), and SNV-BPNN achieved the best prediction effect for total tanshinones (RMSEP = 0.759, RPD = 4.01). When using the SPA and VISSA, the number of wavelengths was reduced below 60 and 150, respectively, and the performance of the models was all very good (RPD > 3). Therefore, the combination of HSI with chemometrics provides a promising method for predicting the active ingredients of Salvia miltiorrhiza and identifying its geographical origins.

Keywords: Salvia miltiorrhiza; chemometrics; content prediction; hyperspectral imaging; traceability.

Figure 1

The specific analysis workflow.

Figure 2

Spectral curves of S. miltiorrhiza powder samples of different origins. (A) The original spectral curve of 420 samples. Each line with a kind of different color represents a wavelength. And (B) the average spectral curves of S. miltiorrhiza powder samples from the five origins.

Figure 3

The confusion matrices on the prediction sets of the PLS-DA and SVM models. (A) D1-PLS-DA, (B) D2-SVM.

Figure 4

The best-performing prediction model of the four chemical indicators based on the full wavelength. (A) Prediction of Tan I by the MSC-PLSR, (B) prediction of Tan IIA by the ORI-BPNN, (C) prediction of CTS by the SG-BPNN, and (D) prediction of the total tanshinones by the SNV-BPNN.

Figure 5

Selected wavelengths for best classification model of S. miltiorrhiza by SPA and VISSA. (A) SPA-D1-PLS-DA, (B) VISSA-D1-PLS-DA.

Figure 6

Selected wavelengths for best prediction of the chemical indicators of S. miltiorrhiza samples by the SPA. The selected wavelengths are shown in red boxes. (A) The elected wavelengths for the prediction of Tan I (MSC-PLSR), (B) the selected wavelengths for the prediction of Tan IIA (ORI-BPNN), (C) the selected wavelengths for the prediction of CTS (SG-BPNN), and (D) the selected wavelengths for the prediction of total tanshinones (SNV-BPNN).