Next Article in Journal
Risk Classification of Food Incidents Using a Risk Evaluation Matrix for Use in Artificial Intelligence-Supported Risk Identification
Previous Article in Journal
A New Strategy for Enhancing Postharvest Quality of Sweet Cherry: High-Voltage Electrostatic Field Improves the Physicochemical Properties and Fungal Community
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Foods
Volume 13
Issue 22
10.3390/foods13223673
foods-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

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

by
Yaoyao Dai
1,2,3,
Binbin Yan
1,3,4,
Feng Xiong
1,3,
Ruibin Bai
1,3,4,
Siman Wang
1,3,
Lanping Guo
1,2,3 and
Jian Yang
1,3,4,*
1
State Key Laboratory for Quality Ensurance and Sustainable Use of Dao-di Herbs, National Resource Center for Chinese Materia Medica, China Academy of Chinese Medical Sciences, Beijing 100700, China
2
School of Pharmacy, Chengdu University of Traditional Chinese Medicine, Chengdu 610072, China
3
Key Laboratory of Biology and Cultivation of Herb Medicine, Ministry of Agriculture and Rural Affairs, Beijing 100700, China
4
Dexing Research and Training Center of Chinese Medical Sciences, Dexing 334213, China
*
Author to whom correspondence should be addressed.
Foods 2024, 13(22), 3673; https://doi.org/10.3390/foods13223673
Submission received: 11 October 2024 / Revised: 1 November 2024 / Accepted: 13 November 2024 / Published: 18 November 2024
(This article belongs to the Section Food Analytical Methods)
Browse Figures
Versions Notes

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.

1. Introduction

Salvia miltiorrhiza Bunge is an herb of the genus Salvia, family Lamiaceae, which is widely distributed in China, such as in provinces of Shandong, Sichuan, Anhui, Hebei, and Shanxi [1,2]. Its dried roots and rhizomes, called “Danshen” in China, are a well-known traditional Chinese medicine, with about 80 million tons used annually. It is commonly used to treat chest paralysis, heart pain, and menstrual disorders. The powder of S. miltiorrhiza is also a common health food used in soups, tea, or beverages [3]. Lipid-soluble tanshinones such as tanshinone I (Tan I), tanshinone IIA (Tan IIA), and cryptotanshinone (CTS) are the main active compounds of S. miltiorrhiza [4], which had been indicated to have antitumor [5], osteoporosis-inhibiting [6], anti-inflammatory [7], anti-anxiety [8], and antithrombotic [9] effects. Furthermore, the sum of Tan I, Tan IIA, and CTS was employed as an important indicator for evaluating the quality of S. miltiorrhiza [10]. The quality of S. miltiorrhiza varied among the geographical origins due to the different environmental conditions [2]. For example, S. miltiorrhiza from Shandong has a higher content of tanshinones [11]. This has led to price differences in S. miltiorrhiza from different origins and sometimes to fraud by sellers claiming false areas of origin. Therefore, a rapid method to access the active compound’s content and distinguish the geographical origin of S. miltiorrhiza is urgently needed.
The traditional methods for determining tanshinones in S. miltiorrhiza include high-performance liquid chromatography (HPLC) [12], ultra-performance liquid chromatography-mass spectrometry/mass spectrometry (UPLC-MS/MS) [13], and ultra-performance liquid chromatography-triple quadrupole mass spectrometry (UPLC-QQQ-MS) [14]. The traditional methods for identifying the origins of herbs include inductively coupled plasma mass spectrometry (ICP-MS) [15], isotope ratio mass spectrometry (IRMS) [16], HPLC [17], etc. Although these methods have the advantages of high accuracy, they also have the disadvantages of being destructive, time-consuming, and costly [18]. Therefore, the quality evaluation and origin tracing of S. miltiorrhiza require a non-destructive, fast, and green technical method.
Hyperspectral imaging (HSI) technology has the advantages of being fast, non-destructive, and green [19]. It can be applied to content prediction, origin tracing, and authenticity identification of traditional Chinese medicines [20]. Compared with infrared and ultraviolet imaging, HSI adds one-dimensional spectral data to normal two-dimensional spatial images to form three-dimensional data [21]. This results in massive hyperspectral data and a large amount of redundant information. HSI combined with chemometrics can provide a more comprehensive quality evaluation and origin identification of traditional Chinese medicine. Liu et al. [22] combined the HSI technique with partial least square regression (PLSR), support vector regression (SVR), and random forest regression (RFR) algorithms to establish a quantitative model. The model could predict the content of Lycium barbarum polysaccharide, total flavonoid, betaine, and vitamin C in Lycium barbarum. Wang et al. predicted the content of ginsenoside Rg2 in ginseng by combining HSI and chemometrics [23]. Liu et al., identified the origin of hawthorn based on hyperspectral [24]. However, there is limited research on origin identification and content prediction of tanshinones in S. miltiorrhiza using HSI.
This study aimed to combine HSI with chemometrics to predict the content of tanshinones in S. miltiorrhiza and to identify their geographical origins. The specific objectives were (1) to investigate the differences in the active compounds in S. miltiorrhiza and to identify S. miltiorrhiza from different origins using HSI in combination with chemometrics; (2) to predict the contents of CTS, Tan IIA, and Tan I using HSI and compare the noise reduction methods and the regression models; and (3) to improve the performance of the model by selecting the wavelengths of HSI using the successive projections algorithm (SPA) and variable iterative space shrinkage approach (VISSA).

2. Materials and Methods

2.1. Sample Collection and Pretreatment

Samples of S. miltiorrhiza were collected from October to December 2021 from thirty cultivated lands in five provinces (Shandong, Hebei, Shanxi, Sichuan, and Anhui) in China (Supplementary Table S1). Fourteen healthy and uniformly growing plants were randomly dug in each cultivated land using the checkerboard sampling method (84 samples per province, 420 samples in total). All samples were identified as the dried roots and rhizomes of S. miltiorrhiza Bge. of the family Lamiaceae by Associate Researcher Yang Jian from the National Resource Center for Chinese Materia Medica, China Academy of Chinese Medical Sciences. The roots and rhizomes of the plants were air-dried and ground into fine powder. Then, the powder samples were sieved through a 65-mesh sieve for spectral acquisition.

2.2. Data Collection Using an HSI System and Extraction of Interest Origins

Spectral information of 420 samples of S. miltiorrhiza was obtained using a visible and short-wave/long-wave near-infrared hyperspectral imaging spectroscopy system (VIS-NIR-HSI, HySpex VNIR-1800/HySpex SWIR 384, Norsk Elektro Optikk, Oslo, Norway). The system consisted of visible near-infrared (VNIR) and short-wave/long-wave near-infrared (SWIR) lenses, two 150 W tungsten halogen lamps (H-LAM, Norsk Elektro Optikk, Oslo, Norway), a conveyor belt, and a computer for data collection. The VNIR spectral range was 410~990 nm, each wavelength interval was 5.4 nm, and a total of 108 wavelengths were obtained. The SWIR spectral range was 950~2500 nm, each wavelength interval was 5.45 nm, and a total of 288 wavelengths were obtained. Then, 2 g of S. miltiorrhiza powder was placed in a 3.5 cm disposable Petri dish, which was shaken gently until the surface was evenly distributed. The exposure times for VNIR and SWIR were 0.0035 s and 0.0045 s, respectively. The samples were placed 30 cm from the lens. The speed of the conveyor belt was 2.5 mm/s.
To attenuate the noise interference due to the dark current and the uneven distribution of the light source in the HSI acquisition process, black and white plate correction of the HSI data was performed before further analysis. The correction formula is as follows:
R = R r   −   R b R w   −   R b
where R denotes the corrected spectral data, Rr denotes the original spectral data, Rb represents the blackboard reference data obtained by turning off the light and blocking the camera lenses, and Rw refers to the whiteboard reference data obtained from the whiteboard with a reflectivity of 99% [25]. The spectral information of each powder sample of S. miltiorrhiza was treated as one origin of interest (ROI) and extracted using the ENVI 5.3 software (Harris Geospatial Solutions Inc., Boulder, CO, USA).

2.3. Chemometrics Analysis

2.3.1. Pretreatment

Spectral data contain not only information related to the composition of the sample but also interference information such as electrical noise, matrix background, and scattering effects. To reduce the influence of interference information and improve the accuracy and reliability of the models, the original spectrum (ORI) was preprocessed in this study using five methods, namely derivative (D1), second derivative (D2), Savitzky–Golay filtering (SG), multiplicative scatter correction (MSC), and standard normal variate transformation (SNV).
Specifically, D1 displays the rate of change in light intensity as a function of wavelength. D2 displays local slope information for spectral data. They could correct the baseline and highlight the peak and absorption features of the spectral data [26]. The SG is commonly employed to eliminate noise. It smooths the data by fitting a polynomial over the data and calculating the value at each point [27]. MSC is widely used to correct scattering effects in spectral data to reflect the true chemical information of the sample [28]. Additionally, SNV is commonly utilized to correct additive and multiplicative information in spectral data and thus improve the interpretability of spectral data [29].

2.3.2. Classification Models

Two methods, partial least squares discriminant analysis (PLS-DA) and support vector machines (SVMs), were used in this study to establish the classification models.
PLS-DA is a supervised linear algorithm developed based on the least squares algorithm, and it is commonly used as a qualitative analysis method in spectral analysis. It establishes a qualitative analysis model based on the characteristic parameters of the known sample set [18], and then it can judge unknown samples. In this study, the characteristic parameters were set at 4~15.
The SVM is a classical supervised learning algorithm. Its core idea is to find a hyperplane in feature space that makes the distance between the samples on either side of the plane as large as possible. So, the generalization of the model is improved. The algorithm can be applied to non-linear classification, high-dimensional data mining, and regression fitting. Also, it is suitable for the classification of small- and medium-sized complex datasets [30].

2.3.3. Prediction Models

Three methods, the back-propagation neural network (BPNN), PLSR, and RF, were employed in this study to establish the content prediction models.
The BPNN is a multilayer feedforward neural network trained by an error back-propagation algorithm. It consists of the input layer, hidden layer, and output layer. Each layer contains multiple nodes. The number of neuron nodes in the input and output layers is determined. The number of neuron nodes in the hidden layer can be changed, and this has a large impact on the prediction results [31].
PLSR is a linear regression method which integrates the advantages of principal component analysis, typical correlation analysis, and multiple linear regression. It can be utilized to solve the problem of multicollinearity of variables and the problem of predicting variables with small sample sizes [32].
RF is an integrated learning algorithm that can be employed to solve classification, regression, and clustering problems for large sample sizes. It consists of classification decision trees. To make a prediction a sample, the random forest obtains the final result by averaging or voting on the predicted values of all the decision trees [33].

2.3.4. Effective Wavelength Screening Algorithms

HSI with many bands and high data dimensions may lead to redundant interference problems, thus increasing the computational burden of modeling. Screening effective wavelengths is a common method to simplify the models and improve the calculation speed of the models. In this study, the SPA and VISSA were adopted to select the effective wavelengths.
The SPA is a forward feature variable selection method. It uses vector projection analysis and correction models to select effective wavelengths with low redundancy and low covariance that can reflect the key information of the sample spectrum [34].
The VISSA is a new algorithm for screening effective wavelengths based on model cluster analysis strategy (MCA) and weighted binary matrix sampling (WBMS). It can eliminate irrelevant variables in each iteration while preserving the key variables, thereby obtaining the optimal combination of variables [20].

2.4. Reference Determination of Four Tanshinone Content

The total tanshinones were the sum of CTS, Tan IIA, and Tan I content. In this study, according to the established method, which was based on the previous published method in our study group [35], CTS, Tan IIA, and Tan I contents were determined by ultra-performance liquid chromatography-quadrupole/trap-mass spectrometry (UPLC-QTRAP-MS/MS-6500, ABSCIEX, Framingham, MA, USA). Specifically, 15 mg of sample powder was placed into a 2 mL centrifuge tube and extracted with 1.5 mL of methanol solution (LC/MS grade, Fisher Scientific, Hampton, NH, USA) for 40 min using an ultrasonic bath (250 W, 40 kHz). Then, centrifugation (12,000 rpm) was performed for 5 min, and the supernatant was obtained through a 0.22 µm filter membrane. The chromatographic column was an ACQUITY UPLC BEH C18 column (2.1 mm × 100 mm, 1.7 µm, Waters, Milford, MA, USA) with a column temperature of 40 °C, an injection volume of 1 µL, and a flow rate of 0.5 mL·min−1. The mobile phases were acetonitrile (LC/MS grade, Fisher Scientific, Hampton, NH, USA) with 0.1% formic acid (LC/MS grade, Fisher Scientific, Hampton, NH, USA) (A) and water with 0.1% formic acid (B). The elution gradients were as follows: 0~3.5 min, 65%~75% A; 3.5~4.5 min, 75%~65% A; and 4.5~6.5 min, 65% A.
Mass spectrometry analyses were conducted in the following conditions. The electrospray ionization source (ESI) mode was configured with positive ion scanning and an ion source temperature of 550 °C. The curtain air (CUR) flow rate was 30 L/min. The spray voltage was 5500 V. The atomizing gas flow rate was 50 L/min. The heating assist gas flow rate was 50 L/min, and the acquisition method was the multi-response monitoring mode (MRM). The conditions of Q1 MASS, Q3 MASS, depolymerization voltage, collision energy, and injection voltage were optimized using standard samples of CTS (CAS: 20090909; Beit Renkang, Beijing, China), Tan IIA (CAS: 22012016; Beit Renkang, Beijing, China), and Tan I (CAS: 20090901; Beit Renkang, Beijing, China). A quantitative regression equation was formulated based on the peak area (Y) and the corresponding concentration (x) of the standard samples. The coefficient of determination was 0.992 5~0.999 9. In the precision test, stability test, repeatability test and recovery test, the RSD of CTS, Tan IIA, and Tan I was 0.64%~2.70%, 2.60%~3.90%, 1.30%~3.30%, and 1.60%~11.00%. The total ion current diagrams of a Salvia miltiorrhiza sample are in Supplementary Figure S1.

2.5. Statistical Analysis

Significant differences (p < 0.05) in tanshinones content of S. miltiorrhiza from different origins were analyzed using the R software (version 4.2.0) with Kruskal–Wallis by adopting the “Games–Howell” multiple comparison method (“userfriendlyscience” package), considering the non-normality of the data. Meanwhile, MATLAB R2024a was used for establishing models and drawing figures.
In the classification model, 420 samples were divided into a calibration set and a prediction set at a ratio of 7.5:2.5. In the prediction model, the samples were divided into a calibration set and a prediction set at a ratio of 7:3. In this study, curve correlation coefficient (R2), root mean squared error (RMSE), and residual predictive deviation (RPD) were taken to evaluate the regression model performance. Specifically, accuracy was used to evaluate the performance of the classification model. When the R2 value ranged between 0.60~0.80 and the RPD value ranged between 2.0~2.5, it indicated that the model could be used for indicator prediction. When the R2 value ranged between 0.80~0.90 and the RPD value ranged between 2.5~3.0, it indicated that the model had a good prediction effect. When the R2 value exceeded 0.90 and the RPD value exceeded 3.0, it indicated that the model had a very good prediction effect [23]. The specific analysis workflow of the study is illustrated in Figure 1.

3. Results

3.1. Statistical Analysis of Tanshinone Content in S. miltiorrhiza from Different Origins

The contents of the four types of tanshinones in S. miltiorrhiza from different origins are listed in Table 1. The result suggested that there were significant differences (p < 0.05) in the contents of the four types of tanshinones in S. miltiorrhiza of different origins. The content of Tan I ranged from 0.132 mg/g in S. miltiorrhiza from Shanxi to 1.090 mg/g in S. miltiorrhiza from Shandong. The content of Tan IIA ranged from 0.826 mg/g in S. miltiorrhiza from Hebei to 2.807 mg/g in S. miltiorrhiza from Shandong. The content of CTS ranged from 0.451 mg/g in S. miltiorrhiza from Hebei to 3.616 mg/g in S. miltiorrhiza from Shandong. The content of total tanshinone ranged from 1.593 mg/g to 7.469 mg/g, and it was the highest in S. miltiorrhiza from Shandong. Meanwhile, the content of total tanshinones in S. miltiorrhiza from Anhui was high. The content of total tanshinones in S. miltiorrhiza from Hebei was the lowest.

3.2. Raw spectra Characteristics of S. miltiorrhiza

The spectral curves of the S. miltiorrhiza powder of different origins exhibited similar trends, but there were some differences in their reflectance intensities (Figure 2A). This may be due to the commonality of the internal tissues of S. miltiorrhiza from different origins and the differences in the content of the chemical components. As demonstrated in Figure 2B, the spectral curve showed a sharp upward trend at 400~900 nm. In the range of 900~2800 nm, the spectral curve showed a fluctuating downward trend with obvious peaks and troughs.
The peaks at different wavelengths were analyzed. The average spectral reflectance intensities of S. miltiorrhiza powder from different origins exhibited large differences in the range of 400~700 nm in the visible region, which may lead to different colors of the samples. The optimal wavelength range for tanshinones was 1220~2580 nm [36], and the S. miltiorrhiza powder samples from Shandong and Anhui had stronger spectral reflectance in this range. This coincided with the phenomenon that the S. miltiorrhiza from Shandong and Anhui had a higher content of tanshinones. The peaks at 910 nm and 960 nm can be attributed to the second overtone of the O-H stretching mode of water or carbohydrates [37]. The peaks at 1120 nm and 1300 nm were related to the second stretching overtones of C-H [38]. The peak at 1740 nm was related to the secondary octave of cyclohexene methylidene stretching vibration, and the peak at 2220 nm was related to the combined frequency of benzene ring C-C stretching vibration and C-H stretching vibration [39].

3.3. Classification of the Geographical Origins of S. miltiorrhiza Based on Hyperspectral Imaging Full Wavelengths

To remove baseline shifts and correct for scattering effects, the original spectrum (ORI) was processed with five pretreatment methods (D1, D2, SG, MSC, and SNV). Then, PLS-DA and SVM models combined with ORI and the five pretreatment methods were established to discriminate the origins of S. miltiorrhiza. The results of the models are presented in Table 2. Generally, the origins of miltiorrhiza could be discriminated using PLS-DA and SVM models, and the PLS-DA models demonstrated better discrimination ability than the SVM models. In the PLS-DA, the discrimination accuracy on the calibration sets and prediction sets was higher than 95%, indicating that PLS-DA models had good discrimination ability for S. miltiorrhiza from different origins. Among all types of PLS-DA models, the D1-PLS-DA model obtained the highest discrimination accuracy of 98.97% on the prediction sets. Among all types of SVM models, the D2-SVM model obtained the highest discrimination accuracy of 95.24% on the prediction sets.
To further measure the discrimination ability and discrimination accuracy of PLS-DA and SVM models for S. miltiorrhiza of different origins, the optimal models D1-PLS-DA and D2-SVM were selected, and the confusion matrices of their prediction sets were plotted (Figure 3). The results indicated that the D1-PLS-DA model can more accurately discriminate S. miltiorrhiza of different origins than the D2-SVM model. In the PLS-DA model, one S. miltiorrhiza sample from Hebei was misidentified as a sample from Shandong (Figure 3A). In the SVM model, two S. miltiorrhiza samples from Shanxi were misidentified as samples from Shandong, and three S. miltiorrhiza samples from Anhui were misidentified as samples from Hebei (Figure 3B).

3.4. Prediction of Chemical Indicators Based on HSI Full Wavelength

The BPNN, PLSR, and RF were combined with five pretreatment methods (D1, D2, SG, MSC, SNV) and ORI to establish the prediction models. The results of the different models for predicting the four chemical indicators of S. miltiorrhiza are presented in Table 3. In the prediction of Tan I, D1-BPNN and D2-BPNN models demonstrated a slight decrease in performance, and the rest of the BPNN and PLSR models obtained good prediction results. Among them, MSC-PLSR (Figure 4A) achieved the best prediction effect. In the prediction of Tan IIA, the ORI-BPNN and MSC-BPNN models obtained good predictive effects, with the ORI-BPNN (Figure 4B) model demonstrating the best prediction effect. In the prediction of CTS, the ORI-BPNN, SG-BPNN, and MSC-BPNN models showed a very good effect. In particular, the SG-BPNN (Figure 4C) model obtained the best prediction effect among them. In the prediction of total tanshinone content, the BPNN and PLSR models combined with ORI, SNV, MSC, and SG obtained excellent prediction results, and the SNV-BPNN (Figure 4D) model obtained the best prediction effect.
Overall, the BPNN and PLSR models showed better prediction capability, and the RF model showed poor prediction capability. It indicated that the RF model was not suitable for predicting the content of tanshinones in S. miltiorrhiza. Among the four components, total tanshinone content was the best predicted. This was probably because Tan I, Tan IIA, and CTS were structurally similar. The three components may interfere with each other for the final spectral information produced when the content of a single component is predicted.

3.5. Classification and Prediction of Chemical Indicators Based on Selected Wavelengths

3.5.1. Classification of the Geographical Origins of S. miltiorrhiza Based on Selected Wavelengths

To obtain the best effective wavelength screening method and further simplify the models, based on the results of the classification models established with full wavelengths, this study extracted the effective wavelengths using the SPA and VISSA methods for the D1-PLS-DA model with the best classification effect. The results are illustrated in Table 4 and Figure 5.
Compared with the full-wavelength models, the performance of the D1-PLS-DA model was improved when the SPA and VISSA were employed. The accuracy on the training and prediction sets all reached 100%. The number of effective wavelengths extracted by the SPA and VISSA was 51 and 150, respectively. It can be found that the number of wavelengths extracted by the SPA is much smaller than that extracted by the VISSA. Further observing the distribution of effective wavelengths, there are 25 same effective wavelengths (Figure 5A,B). This indicated that it may contain information about differences in S. miltiorrhiza samples from different origins.

3.5.2. Prediction of Chemical Indicators Based on HSI Selected Wavelengths

Based on the content prediction results at full wavelengths, the best-performing model for the prediction of the four components was selected to extract effective wavelengths, and the results are shown in Table 5. Comparing the number of effective wavelengths obtained by the SPA and VISSA, the number of effective wavelengths extracted by the SPA for CTS, Tan IIA, Tan I, and the total tanshinones were 36, 33, 21, and 33, respectively. The number of wavelengths was reduced by 90.91%, 91.67%, 94.70%, and 91.67%, respectively. The number of effective wavelengths extracted by the VISSA were 100, 104, 131, and 111, respectively. It can be seen that the SPA had less wavelength information and could simplify the model more. Then, the performance of each effective wavelength model was further compared. With the SPA and VISSA, the models demonstrated good performance for the prediction of the four components. The R2 values on the prediction sets and the training sets were greater than 0.9, and the RPD values were greater than 3. From the results, the number of effective wavelengths extracted by the SPA was much smaller than that of the VISSA when both contributed to good model performance. This suggested that the SPA was more suitable for extracting effective wavelengths for the prediction of chemical indicators in S. miltiorrhiza.
The VISSA obtained eight of the same effective wavelengths for the prediction of the four components, and they were 480, 540, 1035, 1461, 1466, 1499, 1504, and 1510 nm. In comparison, the SPA did obtain the same effective wavelengths, indicating that the effective wavelengths extracted by the SPA are more representative (Figure 6).

4. Discussion

In this study, the tanshinone content of S. miltiorrhiza from different origins was determined. It was found that the tanshinone content of S. miltiorrhiza from Shandong and Anhui was higher, which was consistent with the results obtained by Wang et al. [35]. S. miltiorrhiza is a widely distributed species in China. The quality of S. miltiorrhiza varies from place to place due to differences in environment, germplasm, and cultivation methods, etc. [40]. Liang et al. found that conditions in terms of soil effective phosphorus and precipitation were more favorable for tanshinone IIA accumulation in S. miltiorrhiza in Mengshan, Shandong [2]. Zhang et al. found the specific geospatial distribution of these active ingredients of S. miltiorrhiza in China, due to the climatic factors [41]. Zhang et al. found that although S. miltiorrhiza from the same region, Ji’an (Jiangxi province), were varied according to soil inconsistence, the climate factors among different regions in a larger spatial scale had more effect on S. miltiorrhiza [42]. Therefore, there were differences in the constituent contents of S. miltiorrhiza in the same provinces, but the differences in the contents of S. miltiorrhiza from different provinces were more significant [43]. The consistency of the quality of S. miltiorrhiza was high in Sichuan [44]. The same findings were made for Coptis [45], Sorghum bicolor (L.) Moench [46], and Curcuma Species Grown [47], whose qualities are also differences in different origins. This implies that there is a great market potential for the use of hyperspectral imaging in origin identification.
Sun et al. combined HSI with a SVM to establish a classification model of S. miltiorrhiza from different origins and obtained better results [48]. In this study, in addition to the SVM model, the PLS-DA model was employed, and it was found that the PLS-DA model performed better than the SVM model. In the classification models, the performance was improved after ORI was preprocessed, except for the SNV-SVM model, whose performance was degraded. This may be because, for SVM models, SNV preprocessing amplifies the noise in the raw data, leading to a degradation in model performance.
Infrared spectroscopy has been used for the determination of tanshinone content in S. miltiorrhiza [49], but there is little research on the use of HSI. In this study, it was found that RF models had poor performance in the prediction of tanshinone content of S. miltiorrhiza, with RPD values all less than 2.5. It indicated that RF models were not suitable for the content prediction of S. miltiorrhiza. Among the BPNN and PLSR models, the best prediction models for the four types of tanshinones were not the same. This was consistent with the results obtained by Wang et al. [20]. Different model combinations have different effects on model performance. Xia et al. found that the CARS-BP-ANN model had the best performance to predict chlorogenic acid, luteoloside, and 3,5-Odicaffeoylquinic acid in chrysanthemum [50]. He et al., found that PLSR models with full or selected wavelengths obtained good performance to predict SO2 residual content in Fritillaria thunbergii Bulbus [51]. Li et al., found that the detrending-SPA-BPNN had good model performance to determine N content in citrus leaves [52]. Liu et al. found that the RFOK model had excellent spatial distribution prediction ability for soil heavy metal pollution, compared with the RF model and ordinary kriging model [53]. In the models of CARS, IVSO, or VISSA combined with the PLSR or SVR, Wang et al., found that CARS-SVR exhibited good model performance [54]. Therefore, corresponding models should be selected to predict the composition of samples. However, this is not conducive to developing a miniaturized hyperspectral device that can predict the chemical content of all Chinese medicine. So, in future research, a unified prediction model will be studied in depth.
Combining the effective wavelengths of the classification and prediction models, the same effective wavelengths for the SPA-D1-PLS-DA to predict CTS content were 1313, 2333, and 2403 nm. The same effective wavelengths for the SPA-D1-PLS-DA to predict Tan I were 935, 2458, and 2474 nm. The same effective wavelengths for the SPA-D1-PLS-DA to predict Tan IIA were 1515, 1749, 2300, and 2431 nm. The same effective wavelength for the SPA-D1-PLS-DA to predict total tanshinones was 2414 nm. This suggested that the four types of tanshinones provided part of the information for distinguishing S. miltiorrhiza from different origins. Different effective wavelengths were analyzed in correlation with the chemical structures. The peak near 1300 nm was related to the second overtone of C-H stretching [55], and the peak within 2300~2400 nm was related to the combined frequency of methyl stretching and bending vibrations [32,56]. The peak near 2500 nm was related to the combination of C-H stretching and bending, and the peak within 1500~1904 nm was related to the second overtone of C=O stretching and the first overtone of C-H stretching [57].

5. Conclusions

The origin and content of tanshinones are important indexes to judge the quality of S. miltiorrhiza. The content of tanshinones in S. miltiorrhiza from different origins was different. In this study, HSI combined with chemometrics was used to quickly and accurately identify the origin of S. miltiorrhiza and predict the content of tanshinone components. The SPA largely simplifies the model in addition to making it perform well.
The overall results indicate that HSI technology has great potential for the origin traceability and quality evaluation of S. miltiorrhiza. Compared with traditional methods, it has the advantages of non-destructive and rapid detection. To further improve the applicability of the model, we will increase the chemical indexes of S. miltiorrhiza and expand the collection sources of S. miltiorrhiza in future studies. It is expected to provide some references for the development of proprietary HSI equipment to achieve origin traceability and quality evaluation of S. miltiorrhiza.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/foods13223673/s1, Table S1: Information for the sampling locations of Salvia miltiorrhiza, Figure S1: Total ion current diagrams of a Salvia miltiorrhiza sample: 1: tanshinone I 2: cryptotanshinone, 3: tanshinone IIA.

Author Contributions

Y.D.: Conceptualization, Data curation, Writing—original draft. B.Y.: Writing—review and editing, Project administration, Supervision. F.X.: Writing—review and editing, Data curation. R.B.: Writing—review and editing. S.W.: Methodology, Formal analysis. L.G.: Funding acquisition, Project administration, Supervision. J.Y.: Funding acquisition, Project administration, Supervision. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Shandong Provincial Key Research and Development Program (Major Technological Innovation Project) (2021CXGC010508), Scientific and Technological Innovation Project of China Academy of Chinese Medical Sciences (CI2023E002); National Natural Science Foundation of China (82204566), Innovation Team and Talents Cultivation Program of National Administration of Traditional Chinese Medicine (ZYYCXTD-C-202205), China Agriculture Research System of MOF and MARA (CARS-21), and Dexing Research and Training Center of Chinese Medical Sciences (20222CCH45004).

Data Availability Statement

The original contributions presented in the study are included in the article/Supplementary Materials; further inquiries can be directed to the corresponding author.

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References

  1. Deng, A.P.; Guo, L.P.; Zhan, Z.L.; Huang, L.Q. Decipherment of ancient literature about Danshen. Zhongguo Zhong Yao Za Zhi 2016, 41, 4274–4279. [Google Scholar] [CrossRef] [PubMed]
  2. Liang, H.; Kong, Y.; Chen, W.; Wang, X.; Jia, Z.; Dai, Y.; Yang, X. The quality of wild Salvia miltiorrhiza from Dao Di area in China and its correlation with soil parameters and climate factors. Phytochem Anal. 2021, 32, 318–325. [Google Scholar] [CrossRef] [PubMed]
  3. Zhao, W.; Feng, H.; Guo, S.; Han, Y.; Chen, X. Danshenol A inhibits TNF-α-induced expression of intercellular adhesion molecule-1 (ICAM-1) mediated by NOX4 in endothelial cells. Sci. Rep. 2017, 7, 12953. [Google Scholar] [CrossRef] [PubMed]
  4. Li, X.; Wang, J.; Liu, H. Research Progress on Structural Optimization on the Lipid-Soluble Ingredients of Salvia miltiorrhiza. Chin. J. Org. Chem. 2023, 43, 471–490. [Google Scholar] [CrossRef]
  5. Guo, S.; Wang, Z. Salvianolic acid B from Salvia miltiorrhiza bunge: A potential antitumor agent. Front. Pharmacol. 2022, 13, 1042745. [Google Scholar] [CrossRef]
  6. Guo, Y.; Li, Y.; Xue, L.; Severino, R.P.; Gao, S.; Niu, J.; Qin, L.; Zhang, D.; Brömme, D. Salvia miltiorrhiza: An ancient Chinese herbal medicine as a source for anti-osteoporotic drugs. J. Ethnopharmacol. 2014, 155, 1401–1416. [Google Scholar] [CrossRef]
  7. Yuan, J.; Du, H.; Wan, M.X.; Li, Z.; Zhang, Y.X.; Li, D.K.; Zhuang, P.W.; Ju, A.C. Research progress on anti-inflammatory pharmacological action of components of Salvia miltiorrhiza and its preparations. Drug Eval. Res. 2021, 44, 2322–2332. [Google Scholar] [CrossRef]
  8. Lobina, C.; Colombo, G.; Gessa, G.L.; Carai, M.A.M.; Allegrini, P.; Morazzoni, P.; Riva, A. Anxiolytic effect of an extract of Salvia miltiorrhiza roots in rats. J. Chin. Med. Assoc. 2018, 81, 390–397. [Google Scholar] [CrossRef]
  9. Wu, J.J.; Li, S.L.; Huang, Z.Q.; Gao, R.H.; Wang, H.L.; Peng, Y.; Zhang, C.; Zhao, H.G.; Huang, F. Study on Antioxidative Pharmacodynamics of Aqueous Extract of Salvia miltiorrhiza Bge Leaves in Mice. Strait Pharm. J. 2023, 35, 25–32. [Google Scholar]
  10. Chinese Pharmacopoeia Commission. Pharmacopoeia of the People’s Republic of China; China Medical Science Press: Beijing, China, 2020.
  11. Zhan, Z.L.; Deng, A.P.; Peng, H.S.; Zhang, X.B.; Guo, L.P.; Huang, L.Q. Study of genuineness based on changes of ancient herbal origin--taking Astragalus membranaceus and Salvia miltiorrhiza as examples. Zhongguo Zhong Yao Za Zhi 2016, 41, 3202–3208. [Google Scholar] [CrossRef]
  12. Li, Y.; You, G.; Duan, R.; Liu, K.; Wang, D. Simultaneous Determination of Danshensu, Protocatechuic Acid, Protocatechuic Aldehyde, Rosmarinic Acid and Salvianolic Acid B in Three Processing Products of White Flower Salvia miltiorrhiza by HPLC. Asian J. Chem. 2013, 25, 9558–9560. [Google Scholar] [CrossRef]
  13. Zhang, L.G.; Hu, T.T.; Zhang, F.F.; Luan, S.R.; Li, W.; Deng, H.X.; Lan, Z.H.; Luo, X.F.; Wu, Z.X.; Leslaw, M. Analysis of lipophilic components of Salvia miltiorrhiza roots and S. yunnanensis roots by UPLC and LC-MS/MS. Zhongguo Zhong Yao Za Zhi 2019, 44, 1208–1215. [Google Scholar] [CrossRef] [PubMed]
  14. Zhang, X.; Yu, Y.; Cen, Y.; Yang, D.; Qi, Z.; Hou, Z.; Han, S.; Cai, Z.; Liu, K. Bivariate Correlation Analysis of the Chemometric Profiles of Chinese Wild Salvia miltiorrhiza Based on UPLC-Qqq-MS and Antioxidant Activities. Molecules 2018, 23, 538. [Google Scholar] [CrossRef] [PubMed]
  15. Yuan, M.; Gong, S.; Liu, Y.; Li, X.; Li, M.; Zeng, D.; Li, J.; Guo, Y.; Guo, L. Rapid discrimination of the authenticity and geographical origin of bear bile powder using stable isotope ratio and elemental analysis. Anal. Bioanal. Chem. 2023, 415, 345–356. [Google Scholar] [CrossRef]
  16. Li, X.; Chen, X.Z.; Liu, Z.; Jiang, X.; Zhang, D.L. Identification of Honey Adulteration by Isotope Ratio Mass Spectrometry. Mod. Food Sci. Technol. 2013, 29, 867–871. [Google Scholar] [CrossRef]
  17. Wang, Y.; Jiang, K.; Wang, L.; Han, D.; Yin, G.; Wang, J.; Qin, B.; Li, S.; Wang, T. Identification of Salvia species using high-performance liquid chromatography combined with chemical pattern recognition analysis. J. Sep. Sci. 2018, 41, 609–617. [Google Scholar] [CrossRef]
  18. Liu, Y.; Huang, J.; Li, M.; Chen, Y.; Cui, Q.; Lu, C.; Wang, Y.; Li, L.; Xu, Z.; Zhong, Y.; et al. Rapid identification of the green tea geographical origin and processing month based on near-infrared hyperspectral imaging combined with chemometrics. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 2022, 267, 120537. [Google Scholar] [CrossRef]
  19. Wang, Y.; Yang, J.; Yu, S.; Fu, H.; He, S.; Yang, B.; Nan, T.; Yuan, Y.; Huang, L. Prediction of chemical indicators for quality of Zanthoxylum spices from multi-regions using hyperspectral imaging combined with chemometrics. Front. Sustain. Food Syst. 2022, 6, 1036892. [Google Scholar] [CrossRef]
  20. Wang, Y.; Zhang, Y.; Yuan, Y.; Zhao, Y.; Nie, J.; Nan, T.; Huang, L.; Yang, J. Nutrient content prediction and geographical origin identification of red raspberry fruits by combining hyperspectral imaging with chemometrics. Front. Nutr. 2022, 9, 980095. [Google Scholar] [CrossRef]
  21. Gehlot, S.; Ansari, N.; Gupta, A. WTL-I: Mutual Information-Based Wavelet Transform Learning for Hyperspectral Imaging. Front. Signal Process. 2022, 2, 854207. [Google Scholar] [CrossRef]
  22. Liu, L.L.; Wang, Y.Y.; Yang, J.; Zhang, X.B. Rapid detection technology of chemical component content in Lycii Fructus based on hyperspectral technology. China J. Chin. Mater. Medica 2023, 48, 4328–4336. [Google Scholar] [CrossRef]
  23. Wang, Y.; Zhou, C.; Wang, S.; Yuan, Y.; Bai, R.; Nan, T.; Yang, J. Ginsenoside Rg2 content prediction in Panax ginseng based on the fusion of hyperspectral wavelengths combined with chemometric analysis. J. Food Compos. Anal. 2023, 123, 105619. [Google Scholar] [CrossRef]
  24. Zijian, L.; Jiacheng, G.; Cong, Z.; Youyou, W.; Jian, Y.; Jun, H.; Hongpeng, W.; Ruibin, B. Identification of Geographical Origin for Hawthorn Based on Hyperspectral Imaging Technology. Sci. Technol. Food Ind. 2023, 45, 282–291. [Google Scholar] [CrossRef]
  25. Xuan, G.; Gao, C.; Shao, Y. Spectral and image analysis of hyperspectral data for internal and external quality assessment of peach fruit. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 2022, 272, 121016. [Google Scholar] [CrossRef]
  26. Liu, Y.L.; Lyu, Q.; He, S.L.; Yi, S.L.; Liu, X.F.; Xie, R.J.; Zheng, Y.Q.; Deng, L. Prediction of nitrogen and phosphorus contents in citrus leaves based on hyperspectral imaging. Int. J. Agric. Biol. Eng. 2015, 8, 80–88. [Google Scholar] [CrossRef]
  27. Zulfiqar, M.; Ahmad, M.; Sohaib, A.; Mazzara, M.; Distefano, S. Hyperspectral Imaging for Bloodstain Identification. Sensors 2021, 21, 3045. [Google Scholar] [CrossRef]
  28. Ning, H.; Wang, J.; Jiang, H.; Chen, Q. Quantitative detection of zearalenone in wheat grains based on near-infrared spectroscopy. Spectrochim. Acta A Mol. Biomol. Spectrosc. 2022, 280, 121545. [Google Scholar] [CrossRef]
  29. Xu, J.; Sun, D. Identification of freezer burn on frozen salmon surface using hyperspectral imaging and computer vision combined with machine learning algorithm. Int. J. Refrig. 2017, 74, 151–164. [Google Scholar] [CrossRef]
  30. Chaudhary, S.; Ninsawat, S.; Nakamura, T. Non-Destructive Trace Detection of Explosives Using Pushbroom Scanning Hyperspectral Imaging System. Sensors 2019, 19, 97. [Google Scholar] [CrossRef]
  31. Feng, J.; Liu, Y.; Shi, X.; Wang, Q. Potential of hyperspectral imaging for rapid identification of true and false honeysuckle tea leaves. J. Food Meas. Charact. 2018, 12, 2184–2192. [Google Scholar] [CrossRef]
  32. Ong, P.; Chen, S.; Tsai, C.; Chuang, Y. Prediction of tea theanine content using near-infrared spectroscopy and flower pollination algorithm. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 2021, 255, 119657. [Google Scholar] [CrossRef] [PubMed]
  33. Gong, A.; Zhu, S.; He, Y.; Zhang, C. Grading of Chinese Cantonese Sausage Using Hyperspectral Imaging Combined with Chemometric Methods. Sensors 2017, 17, 1706. [Google Scholar] [CrossRef] [PubMed]
  34. Zhao, X.; Xu, M.; Zhang, W.; Liu, G.; Tong, L. Identification of zinc pollution in rice plants based on two characteristic variables. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 2021, 261, 120043. [Google Scholar] [CrossRef] [PubMed]
  35. Wang, Q.; Yu, Y.J.; Fu, H.Y.; Li, X.; Yang, J. Quality Difference Analysis of Salviae Miltiorrhizae Radix et Rhizoma from Different Origins Based on Multi-index Content Determination Combined with Chemometrics. J. Instrum. Anal. 2023, 42, 389–401. [Google Scholar] [CrossRef]
  36. Li, Z.; Zhou, L.H.; Yang, G.J.; Ye, Z.L. Rapid Determination of Multi-maker Ingredients in Salvia miltiorrhiza by Near Infrared Diffused Reflection Spectroscopy. China Pharm. 2017, 28, 4247–4251. [Google Scholar] [CrossRef]
  37. Yuan, W.D.; Jiang, H.Z.; Yang, S.Y.; Zhang, C.; Zhou, Y.; Zhou, H.P. Geographical Origin Identification of Ningxia Lycium Barbarum Using Hyperspectral Imaging Technology. Food Sci. 2024, 45, 254–260. [Google Scholar] [CrossRef]
  38. Wang, Y.; Xiong, F.; Zhang, Y.; Wang, S.; Yuan, Y.; Lu, C.; Nie, J.; Nan, T.; Yang, B.; Huang, L.; et al. Application of hyperspectral imaging assisted with integrated deep learning approaches in identifying geographical origins and predicting nutrient contents of Coix seeds. Food Chem. 2023, 404, 134503. [Google Scholar] [CrossRef]
  39. Peng, Y.F.; Wang, J.; Wu, Z.S.; Liu, X.N.; Qiao, Y.J. NIR Band Assignment of Tanshinone IIA and Cryptotanshinone by 2D-COS Technology and Model Application Tanshinone Extract. Spectrosc. Spect. Anal. 2022, 42, 1781–1785. [Google Scholar] [CrossRef]
  40. Fang, W.T.; Deng, A.P.; Ren, Z.L.; Nan, T.G.; Kang, L.P.; Guo, L.P.; Huang, L.Q.; Zhan, Z.L. Research progress on quality evaluation of Salviae Miltiorrhizae Radix et Rhizoma (Danshen). Zhongguo Zhong Yao Za Zhi 2018, 43, 1077–1085. [Google Scholar] [CrossRef]
  41. Zhang, C.; Yang, D.; Liang, Z.; Liu, J.; Yan, K.; Zhu, Y.; Yang, S. Climatic factors control the geospatial distribution of active ingredients in Salvia miltiorrhiza Bunge in China. Sci. Rep. 2019, 9, 904. [Google Scholar] [CrossRef]
  42. Zhang, X.D.; Yu, Y.G.; Yang, D.F.; Qi, Z.C.; Liu, R.Z.; Deng, F.T.; Cai, Z.X.; Li, Y.; Sun, Y.F. Chemotaxonomic variation in secondary metabolites contents and their correlation between environmental factors in Salvia miltiorrhiza Bunge from natural habitat of China. Ind. Crop Prod. 2018, 113, 335–347. [Google Scholar] [CrossRef]
  43. Pan, X.M.; Wei, H.; Liu, Y.; Liu, S.X.; Zhang, T.J.; Ma, X.W.; Han, F.N. Study on quality of Salvia miltiorrhiza from different habitats. Chin. Tradit. Herb. Drugs 2011, 42, 1833–1836. [Google Scholar]
  44. Wu, R.Q.; Xie, Y.; Zhang, Y.Q.; Chen, X.F.; Wu, X.Q.; Gao, Y.M. Study on the Quality Consistency of Salvia miltiorrhiza for Different Districts from Sichuan Province. Chin. J. Ethmed. Ethnopharacy 2023, 32, 27–33. [Google Scholar]
  45. Li, J.; Fan, G.; He, Y. Predicting the current and future distribution of three Coptis herbs in China under climate change conditions, using the MaxEnt model and chemical analysis. Sci. Total Environ. 2020, 698, 134141. [Google Scholar] [CrossRef] [PubMed]
  46. Uchimiya, M. Aromaticity of secondary products as the marker for sweet sorghum [Sorghum bicolor (L.) Moench] genotype and environment effects. J. Agric. Food Res. 2022, 9, 100338. [Google Scholar] [CrossRef]
  47. Lee, J.; Jung, Y.; Shin, J.; Kim, H.K.; Moon, B.C.; Ryu, D.H.; Hwang, G. Secondary Metabolite Profiling of Curcuma Species Grown at Different Locations Using GC/TOF and UPLC/Q-TOF MS. Molecules 2014, 19, 9535–9551. [Google Scholar] [CrossRef]
  48. Sun, Y.C.; Jiao, L. Identification of Salvia miltiorrhiza Regions by Hyperspectrum and Support Vector Machine. Fujian Anal. Test. 2023, 32, 11–15. [Google Scholar] [CrossRef]
  49. Zhang, S.S.; Wang, D.Q.; Zhu, J.J.; Li, C.Y.; Wang, Z.M.; Kuang, Y.H. Process Quality Control of Tanshinones in Salvia miltiorrhiza with the extraction and concertration of Fufangdanshen Tabletsby Near Infrared Spectroscopy. Ginsheng Res. 2021, 33, 15–20. [Google Scholar] [CrossRef]
  50. Xia, Z.; Sun, Y.; Cai, C.; He, Y.; Nie, P. Rapid Determination of Chlorogenic Acid, Luteoloside and 3,5-O-dicaffeoylquinic Acid in Chrysanthemum Using Near-Infrared Spectroscopy. Sensors 2019, 19, 1981. [Google Scholar] [CrossRef]
  51. He, J.; Zhang, C.; He, Y. Application of Near-Infrared Hyperspectral Imaging to Detect Sulfur Dioxide Residual in the Fritillaria thunbergii Bulbus Treated by Sulfur Fumigation. Appl. Sci. 2017, 7, 77. [Google Scholar] [CrossRef]
  52. Lu, Z.; Yu, H.; Yin, Y.; Yuan, Y.; Liang, H.; Li, F.; Li, Z. Determination of the Acid and Peroxide Values of Vegetable Oils by Raman Spectroscopy with Competitive Adaptive Reweighted Sampling (CARS) and Back Propagation Neural Network (BPNN). Anal. Lett. 2024, 57, 2289–2306. [Google Scholar] [CrossRef]
  53. Liu, G.; Zhou, X.; Li, Q.; Shi, Y.; Guo, G.; Zhao, L.; Wang, J.; Su, Y.; Zhang, C. Spatial distribution prediction of soil As in a large-scale arsenic slag contaminated site based on an integrated model and multi-source environmental data. Environ. Pollut. 2020, 267, 115631. [Google Scholar] [CrossRef] [PubMed]
  54. Wang, F.; Wang, C.; Song, S.; Xie, S.; Kang, F. Study on starch content detection and visualization of potato based on hyperspectral imaging. Food Sci. Nutr. 2021, 9, 4420–4430. [Google Scholar] [CrossRef] [PubMed]
  55. Zhang, Y.; Wang, Y.Y.; Zhang, T.; Qin, Y.C.; Nan, T.G.; Yang, J.; Lu, Y.J. Identification of Citri Reticulatae Pericarpium Form Different Scales Geographical Origin by Hyperspectral Imaging Combined with Image Segmentation Algorithm. Chem. Reag. 2023, 45, 136–143. [Google Scholar] [CrossRef]
  56. Xiao, D.; Wang, S.M.; Zhang, Y.; Liu, D.F.; Hao, Q.X.; Bai, R.B.; Yang, J. Research on the detection of andrographolide components in Andrographis paniculata based on hyperspectral technology. Chem. Reag. 2024, 46, 89–98. [Google Scholar] [CrossRef]
  57. Haruna, S.A.; Li, H.; Wei, W.; Geng, W.; Luo, X.; Zareef, M.; Yao-Say Solomon Adade, S.; Ivane, N.M.A.; Isa, A.; Chen, Q. Simultaneous quantification of total flavonoids and phenolic content in raw peanut seeds via NIR spectroscopy coupled with integrated algorithms. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 2023, 285, 121854. [Google Scholar] [CrossRef]
Foods 13 03673 g001
Figure 1. The specific analysis workflow.
Figure 1. The specific analysis workflow.
Foods 13 03673 g001
Foods 13 03673 g002
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 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.
Foods 13 03673 g002
Foods 13 03673 g003
Figure 3. The confusion matrices on the prediction sets of the PLS-DA and SVM models. (A) D1-PLS-DA, (B) D2-SVM.
Figure 3. The confusion matrices on the prediction sets of the PLS-DA and SVM models. (A) D1-PLS-DA, (B) D2-SVM.
Foods 13 03673 g003
Foods 13 03673 g004
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 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.
Foods 13 03673 g004
Foods 13 03673 g005
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 5. Selected wavelengths for best classification model of S. miltiorrhiza by SPA and VISSA. (A) SPA-D1-PLS-DA, (B) VISSA-D1-PLS-DA.
Foods 13 03673 g005
Foods 13 03673 g006
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).
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).
Foods 13 03673 g006
Table 1. Content of tanshinones in S. miltiorrhiza of different origins.
Table 1. Content of tanshinones in S. miltiorrhiza of different origins.
Collection OriginsTan I (mg/g)Tan IIA (mg/g)CTS (mg/g)Total (mg/g)
SD (n = 84)1.090 ± 0.373 d2.807 ± 0.983 d3.616 ± 2.070 d7.469 ± 3.126 d
HB (n = 84)0.316 ± 0.268 ab0.826 ± 0.633 a0.451 ± 0.379 a1.593 ± 1.231 a
SX (n = 84)0.132 ± 0.058 c1.142 ± 0.608 b0.640 ± 0.420 b1.914 ± 1.014 a
SC (n = 84)0.271 ± 0.186 a1.663 ± 1.317 c0.678 ± 0.687 ab2.612 ± 2.066 b
AH (n = 84)0.376 ± 0.257 b1.940 ± 1.113 c1.928 ± 1.831 c4.245 ± 2.786 c
The five origins include Shandong (SD), Hebei (HB), Shanxi (SX), Sichuan (SC), and Anhui (AH). Lowercase letters such as a, b, c, etc., indicate a statistical difference (p < 0.05) between different letters and no statistical difference between the same letters.
Table 2. Discrimination of S. miltiorrhiza origins using PLS-DA and SVM models.
Table 2. Discrimination of S. miltiorrhiza origins using PLS-DA and SVM models.
PretreatmentsPLS-DA SVM
Calibration Set (%)Prediction Set (%)Calibration Set (%)Prediction Set (%)
ORI98.4595.8898.7380.95
D199.6998.9799.3791.43
D298.4597.9499.0595.24
SG99.0797.9498.7380.95
MSC99.0796.9197.4685.71
SNV99.6997.94100.0075.24
Table 3. Four chemical indicators prediction based on the full wavelengths.
Table 3. Four chemical indicators prediction based on the full wavelengths.
Chemical
Indexes		ModelsCalibration SetPrediction SetChemical
Indexes		ModelsCalibration SetPrediction Set
R2RMSECR2RMSEPRPDR2RMSECR2RMSEPRPD
Tan IORI-BPNN0.9650.0790.9240.1183.35Tan IIAORI-BPNN0.9450.2820.9170.3323.34
D1-BPNN0.9480.1080.8610.1632.53D1-BPNN0.9490.2750.8710.4312.63
D2-BPNN0.9500.0950.8460.1722.51D2-BPNN0.8990.3840.8200.5032.33
SG-BPNN0.9660.0790.9170.1253.18SG-BPNN0.9180.3480.8850.3912.66
MSC-BPNN0.9710.0710.9190.1213.41MSC-BPNN0.9390.2970.9190.3283.28
SNV-BPNN0.9610.0830.9300.1143.44SNV-BPNN0.9570.2580.8860.3932.81
ORI-PLSR0.9600.0810.9340.1173.90ORI-PLSR0.9400.2960.8640.4222.66
D1-PLSR0.9500.0910.9060.1363.24D1-PLSR0.9460.2810.8210.4892.32
D2-PLSR0.9400.1000.9100.1323.25D2-PLSR0.8950.3910.8260.4782.32
SG-PLSR0.9700.0720.9310.1233.79SG-PLSR0.9350.3080.8510.4422.51
MSC-PLSR0.9550.0870.9380.1104.03MSC-PLSR0.9500.2690.8530.4392.53
SNV-PLSR0.9760.0630.9320.1203.83SNV-PLSR0.9320.3160.8410.4552.39
ORI-RF0.96000.1190.8800.2181.64ORI-RF0.9460.4000.8600.5781.78
D1-RF0.97950.0840.9230.1752.16D1-RF0.9740.2910.9190.4482.21
D2-RF0.97610.0990.9280.1732.06D2-RF0.9550.3900.9150.4721.93
SG-RF0.96220.1170.8680.2271.59SG-RF0.9520.3850.8530.5921.73
MSC-RF0.96410.1100.8910.2111.77MSC-RF0.9520.3730.8670.5721.91
SNV-RF0.96900.1020.9020.1962.03SNV-RF0.9590.3540.9020.4852.18
CTSORI-BPNN0.9660.3270.9110.5343.03Total (Tan I + Tan IIA + CTS)ORI-BPNN0.9560.6580.9330.8033.83
D1-BPNN0.8810.6120.8600.6592.56D1-BPNN0.9390.7780.8861.1062.96
D2-BPNN0.8970.5630.7110.9661.68D2-BPNN0.9390.7580.8031.3842.15
SG-BPNN0.9270.4830.9110.5273.25SG-BPNN0.9170.8880.9200.8803.24
MSC-BPNN0.8810.6070.9130.5293.01MSC-BPNN0.9510.6990.9270.8563.41
SNV-BPNN0.9260.4800.9070.5562.79SNV-BPNN0.9430.7940.9400.7594.01
ORI-PLSR0.9270.4550.8300.7832.25ORI-PLSR0.9340.7700.9290.8663.53
D1-PLSR0.9010.5300.7700.9141.92D1-PLSR0.9670.5470.8641.2062.66
D2-PLSR0.8970.5380.7110.9111.75D2-PLSR0.9400.7330.8591.2162.44
SG-PLSR0.9270.5820.9110.7182.39SG-PLSR0.9450.7050.9190.9273.32
MSC-PLSR0.9070.5130.8170.8112.14MSC-PLSR0.9310.7870.9090.9853.02
SNV-PLSR0.9030.5250.8230.7982.17SNV-PLSR0.9490.6750.9070.9863.11
ORI-RF0.9500.5500.8780.9561.50ORI-RF0.9710.7380.9121.3691.99
D1-RF0.9680.4410.8940.8901.63D1-RF0.9770.9480.9061.3742.03
D2-RF0.9600.4940.9040.8421.78D2-RF0.9770.6870.9271.2752.05
SG-RF0.9490.5420.8760.9471.56SG-RF0.9630.8280.9071.4181.89
MSC-RF0.9530.5190.8880.8791.83MSC-RF0.9710.7170.9171.2872.30
SNV-RF0.9580.4950.8720.9301.73SNV-RF0.9700.7480.9131.3432.06
Table 4. The classification results of the D1-PLS-DA model based on effective wavelengths.
Table 4. The classification results of the D1-PLS-DA model based on effective wavelengths.
ModelMethodsLVNumberCalibration Set (%)Prediction Set (%)
D1-PLS-DASPA551100100
VISSA13150100100
Table 5. The prediction of four chemical indicators based on the effective wavelengths selected.
Table 5. The prediction of four chemical indicators based on the effective wavelengths selected.
ChemicalModelMethodWavelengths NumberCalibration SetPrediction Set
R2RMSECR2RMSEPRPD
Tan IMSC-PLSRSPA210.9310.1070.9370.1093.94
VISSA1310.9550.0870.9400.1134.08
Tan IIAORI-BPNNSPA330.9060.3690.9050.3573.17
VISSA1040.9520.2630.9020.3623.12
CTSSG-BPNNSPA360.9620.3450.9100.5283.24
VISSA1000.9310.4760.9020.5603.62
TotalSNV-BPNNSPA330.9410.7590.9330.8303.44
VISSA1110.9590.6260.9310.8303.58
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Dai, Y.; Yan, B.; Xiong, F.; Bai, R.; Wang, S.; Guo, L.; Yang, J. Tanshinone Content Prediction and Geographical Origin Classification of Salvia miltiorrhiza by Combining Hyperspectral Imaging with Chemometrics. Foods 2024, 13, 3673. https://doi.org/10.3390/foods13223673

AMA Style

Dai Y, Yan B, Xiong F, Bai R, Wang S, Guo L, Yang J. Tanshinone Content Prediction and Geographical Origin Classification of Salvia miltiorrhiza by Combining Hyperspectral Imaging with Chemometrics. Foods. 2024; 13(22):3673. https://doi.org/10.3390/foods13223673

Chicago/Turabian Style

Dai, Yaoyao, Binbin Yan, Feng Xiong, Ruibin Bai, Siman Wang, Lanping Guo, and Jian Yang. 2024. "Tanshinone Content Prediction and Geographical Origin Classification of Salvia miltiorrhiza by Combining Hyperspectral Imaging with Chemometrics" Foods 13, no. 22: 3673. https://doi.org/10.3390/foods13223673

APA Style

Dai, Y., Yan, B., Xiong, F., Bai, R., Wang, S., Guo, L., & Yang, J. (2024). Tanshinone Content Prediction and Geographical Origin Classification of Salvia miltiorrhiza by Combining Hyperspectral Imaging with Chemometrics. Foods, 13(22), 3673. https://doi.org/10.3390/foods13223673

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop