• Users Online: 126
  • Print this page
  • Email this page

 
Table of Contents
ORIGINAL ARTICLE
Year : 2020  |  Volume : 6  |  Issue : 4  |  Page : 469-480

Geographical origin and level identification of frankincense based on hyperspectral image


School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China

Date of Submission29-Mar-2020
Date of Acceptance11-Jun-2020
Date of Web Publication05-Oct-2020

Correspondence Address:
Prof. Wei Li
School of Information and Electronics, Beijing Institute of Technology, Beijing 100081
China
Login to access the Email id

Source of Support: None, Conflict of Interest: None


DOI: 10.4103/wjtcm.wjtcm_53_20

Rights and Permissions
  Abstract 


Background: As the demand for traditional Chinese medicinal materials increases in China and even the world, there is an urgent need for an effective and simple identification technology to identify the origin and quality of the latter and ensure the safety of clinical medication. Mineral element analysis and isotope finger-printing are the two commonly used techniques in traditional origin identification. Both of these techniques require the use of stoichiometric methods in the identification process. Although they have high accuracy and sensitivity, they are expensive and inefficient. In addition, near-infrared spectroscopy is a fast, nondestructive, and widely used identification technique developed in recent years, but its identification results are susceptible to samples' states and environmental conditions, and its sensitivity is low. Hyperspectral imaging combines the advantages of imaging technology and optical technology, which can simultaneously access the image information and spectral information which reflect the external characteristics, internal physical structure, and chemical composition of the samples. Hyperspectral imaging is widely applied to agricultural product inspection, but research into its application in origin and quality identification of TCM materials is rare. Methods: In this study, the algorithm framework discriminative marginalized least squares regression (DMLSR) was used for feature extraction of frankincense hyperspectral data. The DMLSR with intraclass compactness graph and manifold regularization can efficiently learn the projective samples with higher separability and less redundant information than the original samples. Then, the discriminative collaborative representation with Tikhonov regularization (DCRT) was applied for classifying the geographical origin and level of frankincense. DCRT introduces the discriminant regularization term and incorporates SID, which is more sensitive to the spectrum as the measurement method and is more suitable for the frankincense spectral data compared with SVM. Results: For the origin classification task, samples of all levels from each origin were, respectively, selected for three-way classification. We used 10-fold cross-validation to select a model parameter in the experiment. When obtaining the optimal parameters, we randomly selected the training set and testing set, where the training set accounts for 70% and the training set for 30%. After repeating this random process 10 times, we obtained the final average classification accuracy, which is higher than 90%, and the standard deviation fluctuation is usually small. For the level classification task, samples of each level from three origins were separately selected for multiclassification. We randomly selected the training set and testing set from each origin. The level classification results of the three origins are good on D4350 data, and the classification accuracy of each level is basically above 80%. Conclusion: Experiments and analysis show that our algorithm framework has excellent classification performance, which is stable in origin classification and has potential for generalization. In addition, the experiments show that in our algorithm framework, different classification tasks need to combine different data sources to achieve better classification and recognition, as the origin classification task uses frankincense's D3000 data, and level classification task uses frankincense's D4350 data.

Keywords: Discriminative collaborative representation with Tikhonov regularization, discriminative marginalized least squares regression, frankincense, geographical origin, hyperspectral


How to cite this article:
Zhang YX, Gao ZC, Liu YX, Li W. Geographical origin and level identification of frankincense based on hyperspectral image. World J Tradit Chin Med 2020;6:469-80

How to cite this URL:
Zhang YX, Gao ZC, Liu YX, Li W. Geographical origin and level identification of frankincense based on hyperspectral image. World J Tradit Chin Med [serial online] 2020 [cited 2021 Jan 24];6:469-80. Available from: https://www.wjtcm.net/text.asp?2020/6/4/469/303583




  Introduction Top


Background

Opinions of the CPC Central Committee and the State Council on Promoting the Inheritance, Innovation and Development of Traditional Chinese Medicine was released on the evening of October 26, 2019. Opinions pointed out that the policy of attaching equal importance to both traditional Chinese medicine (TCM) and Western medicine still needs to be comprehensively implemented, the governance system following the law of TCM needs to be improved, the TCM development foundation and talents training are relatively weak, the quality of traditional Chinese medicinal materials varies, the inheritance and innovation of TCM is insufficient, the role of TCM is not fully maximized, and the law on TCM needs to be implemented urgently in an in-depth manner. Effective measures need to be adopted to resolve the aforementioned problems to effectively inherit, develop, and utilize TCM, a valuable inheritance from ancestors.[1] Opinions is the first TCM document released by the CPC Central Committee and the State Council. The National Conference on TCM, held in Beijing on October 25, 2019, is the first national conference on TCM held by the State Council since the founding of the People's Republic of China. The conference called for inheriting the essence, upholding integrity and innovation, and using innovation to develop TCM. Chinese President Xi stressed that TCM, which is a treasure of the Chinese civilization, comprises thousands of years of health concepts and practical experience of the Chinese people and embodies the vast wisdom of the Chinese nation and its people.[2] In the 70 years since the founding of the People's Republic of China, the party and the government have attached great importance to TCM.

As the demand for traditional Chinese medicinal materials increases in China and even the world, problems such as mixed origin medicines, selling substandard medicine at top-quality prices, and mixing false materials with genuine ones in the TCM market are becoming increasingly common, all of which will lead to serious consequences. There is an urgent need for an effective and simple identification technology in the production chain and market circulation of Chinese medicinal materials to identify the origin and quality of the latter and ensure the safety of clinical medication.

Frankincense is the resin from the bark of olivine frankincense tree, as shown in [Figure 1]. It is warm-natured, bitter in taste, with the functions of promoting circulation of blood and qi, relieving pain, and detoxifying. It is clinically mainly used for the treatment of qi and blood stagnation, abdominal pain, carbuncle sore swelling, poison, injury, dysmenorrhea, postpartum blood stasis,[3] etc., Studies have shown that frankincense mainly contains triterpenoids, which have anti-inflammatory[4] as well as analgesic, immunosuppressive, and antitumor properties.[5] Frankincense is often used in perfumes and aromatherapy, and it is sometimes used as an ingredient in skincare products. In TCM, frankincense has antibacterial and blood-activating effects. In Persian medicine, it is used to treat diabetes.[6]
Figure 1: Frankincense olibanum resin

Click here to view


According to the difference of origin, frankincense is classified into Somalian frankincense and Ethiopian frankincense. There is also frankincense from India, but it has not been part of the pharmacopeia. Each frankincense also has different quality levels because of the harvest environment, time, and other factors. They are often sorted manually. Therefore, it is necessary to introduce hyperspectral imaging technology with its characteristics of rapid, nondestructive, and batch detection into the field of TCM identification, so as to solve the problems of mixed origin and shoddy quality in the circulation of TCM.

Purpose and significance

There has been a lot of research into TCM materials' origins, quality, and production estimates in the field of medicinal botany. Because of the diversity of medicinal plant species, however, and the great difference in structure and composition between them, effective and simple artificial recognition methods are still in the process of continuous exploration and development.

Mineral element analysis and isotope fingerprinting are the two commonly used techniques in traditional origin identification.[7],[8] Both of these techniques require the use of stoichiometric methods in the identification process. Although they have high accuracy and sensitivity, they are expensive and inefficient. In addition, near-infrared spectroscopy is a fast, nondestructive, and widely used identification technique developed in recent years,[9],[10],[11] but its identification results are susceptible to samples' states and environmental conditions, and its sensitivity is low. Hyperspectral imaging combines the advantages of imaging technology and optical technology, which can simultaneously access the image information and spectral information which reflect the external characteristics, internal physical structure, and chemical composition of the samples. Hyperspectral imaging is widely applied to agricultural product inspection, but research into its application in origin and quality identification of TCM materials is rare.

Hyperspectral remote sensing is an imaging and spectral imaging technology, which reflects the difference of diagnostic spectral absorption caused by the electron transition and molecular vibration inside the material. It is an important means of earth observation and an indispensable part of the space information network. The data structure of hyperspectral images is shown in [Figure 2]. Hyperspectral images have extremely high spectral range and spectral resolution, among which the spectral resolution is 5–10 nm and the spectral range is 400–2500 nm, with hundreds of spectral bands. With the development of modern information technology, hyperspectral image technology has been widely used in geological surveys, vegetation detection, urban planning, environmental monitoring, medicine, military, and other fields, but the application of hyperspectral technology in TCM or medicinal plants is nonetheless relatively rare.
Figure 2: Hyperspectral image data structure

Click here to view


As an imaging technology which has a wider frequency band and smaller spectrum resolution compared with general ones, hyperspectral imaging technology can reflect more features of objects, including molecular-level features like reflectance spectral characteristics influenced by electron transitions. Hyperspectral imaging is able to observe both the external characteristics and internal quality of an object, making hyperspectral imaging technology theoretically applicable to material identification and classification, and able to obtain more comprehensive features of the measured objects compared with other technologies, so as to achieve more accurate and efficient identification and classification work. It is necessary to research the application of hyperspectral image technology in the field of TCM.


  Research Contents Top


According to the above analysis of the background and purpose, in this research, we took the TCM material frankincense as the research object to conduct two main research works. First, we extracted the features of the frankincense hyperspectral data provided and select a fast, effective feature extraction algorithm according to the characteristics of high correlation between bands, high information redundancy, and low separability of frankincense spectral data. Second, a classification model suitable for the classification of frankincense hyperspectral data with low computational complexity was selected. Based on the feature extracted, the classification model was trained and tested, and the test results of different classifiers were compared for analysis and evaluation. Finally, a complete algorithm framework was integrated.

In this study, we built an algorithm framework in which a discriminative marginalized least squares regression (DMLSR) was used for feature extraction, and discriminative collaborative representation with Tikhonov (DCRT) regularization was used for classification of the geographical origin and level of frankincense. Through comparison with the traditional classifier, it was concluded that our algorithm framework can effectively realize the purpose of frankincense classification and identification. This research proved the feasibility of origin and quality identification using hyperspectral imaging technology, which provided a new technology reference for TCM material identification, and provided an algorithm framework of automatic frankincense origin and quality identification.


  Research Program Top


The overall research approach is shown in [Figure 3]. The training and testing sets are first extracted from the frankincense HSI data from three different origins (including India, Ethiopia, and Somalia) with three, four, or five different levels, which is performed with ENVI. The regions of interest (ROI) are manually selected in each image for 10–25 which are then averaged. The number of ROIs in each image is shown in [Table 1], in which there are four different levels for frankincense from India (1, 2, 3, and 4), five for Ethiopia (A, 1, 2, 3, and 4), and three for Somalia (1, 2, and 3). Asides from this, there are two sets of data for each type of frankincense, corresponding to two band ranges. The data with 288 bands covering 950 nm to 2500 nm are named D3000, and the data with 108 bands covering from 410 nm to 990 nm are named D4350.
Figure 3: Overall research program block diagram

Click here to view
Table 1: The number of regions of interest of different levels in different regions

Click here to view


Analyzing data from samples of selected ROIs, it can be found that the data of different origins and levels have a poor separability in their original space. In other words, every class overlaps on neighbors so that features are inconspicuous and indistinguishable. The visualization results are shown in [Figure 4]. For the poor separability of data in original space, we proposed the use of the DMLSR[12] algorithm for feature extraction, which learns the potential low-dimensional subspace of data where they have a better separability for the compact relationship among intraclass samples and more separable interclass samples.
Figure 4: Visualization of raw data distribution. (a) Visualization of data from different origins (D3000). (b) Visualization of data in different levels from Ethiopia

Click here to view


According to the richness of data's spectral information, we proposed to use the DCRT regularization to continue the classification tasks after the original data have had the features extracted with the DMLSR algorithm.

Research method

The framework of proposed algorithm is mainly based on DMLSR for feature extraction and the DCRT classifier for classification. The declaration of variables involved is shown in [Table 2]. Detailed description of the above two algorithms is in the following sections.
Table 2: Variable declaration

Click here to view


Discriminative marginalized least squares regression algorithm

The use of class label for least squares regression has attracted wide attention in facial recognition, feature representation, feature extraction, and other pattern recognition fields. DMLSR is an extended algorithm based on the least squares regression. [Figure 5] shows its flowchart, in which training sample fits with target matrix R under the constraints of intraclass compactness graph and data reconstruction, and generates projection matrix Q after iterative training. Then, Q is used for feature extraction, and projection samples XTQ and YTQ of training set and testing set are obtained, respectively.
Figure 5: Flowchart of the discriminative marginalized least squares regression feature extraction

Click here to view


The framework of the DMLSR algorithm is formulated as follows:



where R ϵ Rn × c is the target matrix to fit in the training process holding class label prior, and Tr (QTXLXTQ) is the manifold regular term with class compactness diagram for improving the intraclass compact relationship of projection samples in subspace. L is a Laplacian matrix equivalent to D − W, and D is a diagonal matrix. When Xi and Xj are in the same class, the weight coefficient of weight matrix W is calculated from a Gaussian function where . Otherwise, wij = 0. Moreover, is applied to calculating kernel parameters adaptively which shows the local scaling in the neighborhood of xi, of data samples, reflecting the similarity between central pixel and neighboring pixels. is the marginal constraint of the target matrix, where li shows the real class index of xi. The constraint fixes the distance between real and fake classes, able to widen the edge between classes in projection samples XTQ. X = PQTX is the data reconstruction constraint, where Pis an orthogonal reconstruction matrix, which ensures that the extracted features have the main discriminative information in low-dimensional subspace.

Less-redundant, easier-to-distinguishing features can be extracted with DMLSR, which ensures that the following classification process has a higher accuracy.

Discriminative collaborative representation with Tikhonov classifier

Owing to the rise of compressed sensing, representation-based classification algorithms have attracted a lot of attention. After decades of development, there have been many excellent classification algorithms in the fields of face recognition, computer vision, and hyperspectrum. Li et al. proposed the nearest regularization subspace with distance-weighted Tikhonov regularization,[13] and proposed a Collaborative representation with tikhonov (CRT) regularization,[14] which performs well in hyperspectral classification. CRT introduces the Tikhonov regularization factor on the basis of collaborative representation-based classifier. It determines the class label of different materials by means of linear regression, whose core idea is to ensure that the distance between the testing sample and training sample is as small as possible. Therefore, distance-weighted Tikhonov regularization is adopted to calculate the all available linear combination of the training sample. DCRT is an improved algorithm based on CRT. We introduced a discriminative regularization term:



where is the regularization term. Minimizing this term can make the linear representation of all kinds of training samples have the lowest correlation in the feature space, that is, the linear representation is more discriminative. Y ϵ Rd × 1 is the testing sample, and Xiϵ Rd × n1 is the representation coefficient corresponding to the training sample class l = 1, 2… c. n1 is the number of training samples of class l, α1 . αlis the representation coefficient corresponding to class l. λ and γ are regularization parameters to balance the influence of the residual term and Tikhonov regularization term on the representation coefficients. Γl,y is the Tikhonov matrix corresponding to the class l and testing sample y, which is defined as follows:



where x1, x2…., xn is each row of training sample matrix X. In general, Euclidean distance (ED) is adopted as the similarity measurement strategy between training and testing samples, which means that the diagonal elements of Γl,y are derived from ED.

The representation coefficient αl of DCRT can be solved from a closed-form solution:



where M is defined as:



In addition, considering that the spectral information measurement method is more effective than ED, we incorporated spectral information divergence (SID) into the Tikhonov regularization term. The SID theory comes from the relative entropy in information theory, which is mainly used to describe the difference between two variables, which obtains measurements from the view of spectrum. Compared with ED, SID is more sensitive to the fluctuation of the spectrum itself. Γl,y is defined as:



The DCRT classifier based on representation learning has a low computational complexity. On the basis of extracted features from DMLSR, a more robust and generalized classification model can be trained to perform more efficient classification of origin and levels.


  Experimental Design and Analysis Top


Experimental data

The data of this experiment were provided by the Chinese Medicine Resource Centre, China Academy of Chinese Medical Science. HySpex series of hyperspectral imagers were used to collect 410–2500 nm spectral data of frankincense of multiple levels from India, Ethiopia, and Somalia. These data are hyperspectral raw data (digital number value), which have been calibrated by the device's built-in RAD and obtained by two lenses (410–990 nm and 950–2500 nm) with a spectral resolution of 6 nm. The hyperspectral image data of India level 1 (India [1]), Ethiopia level 4 (Ethiopia [4]), and Somalia level 2 (Somalia [2]) are shown in [Figure 6]. The average value of all ROI samples from each origin is calculated to obtain the spectral curves corresponding to the data of three origins, as shown in [Figure 7].
Figure 6: Frankincense hyperspectral image data. (a) India (1) (D3000), (b) Ethiopia (4) (D3000), (c) Somalia (2) (D3000), (d) India (1) (D4350), (e) Ethiopia (4) (D4350), (f) Somalia (2) (D4350)

Click here to view
Figure 7: Spectral curves of samples from different origins. (a) D3000, (b) D4350

Click here to view


Experimental design

For the origin classification task, samples of all levels from each origin were, respectively, selected for three-way classification. We used 10-fold cross-validation to select a model parameter in the experiment. When obtaining the optimal parameters, we randomly selected the training set and testing set, where the training set accounts for 70% and the training set for 30%. After repeating this random process 10 times, we obtained the final average classification accuracy.

For the level classification task, samples of each level from three origins were separately selected for multiclassification. We also used 10-fold cross-validation to select model parameter in the experiment. When obtaining the optimal parameters, we randomly selected the training set and testing set from each origin, where the training set accounts for 70% and the training set for 30%. After repeating this random process 10 times, we obtained the final average classification accuracy and the standard deviation.

In this experiment, support vector machine (SVM) and CRT classifiers were used as the comparison algorithm, and the classification performance was evaluated using recall rate, average classification accuracy (overall accuracy), and kappa coefficient (KC).


  Results and Analysis Top


Parameter optimization

There are four parameters to be tuned in the algorithm framework of this study, λ1 and λ2 in DMLSR and λ and γ in DCRT, all of which float in a range of{1e − 5, 1e − 4, 1e − 3, 1e − 2, 1e − 1, 1e − 0, 5e + 0, 1e + 1}.

In this experiment, the optimal parameters were selected using the 10-fold cross-validation strategy. Taking origin classification (by D3000) as an example, we first put the features extracted from DMLSR into the K-nearest neighbor classifier to obtain the 10-fold average accuracy curves of λ1 and λ2, as shown in [Figure 8]a, and selected the parameters with the highest corresponding classification accuracy. Second, we fixed the optimal values of λ1 and λ2, extracted the optimal features and put them into DCRT for λ and γ tuning, obtained the 10-fold average precision curves of λ and γ, as shown in [Figure 8]b, and selected the optimal values of λ and γ. The above tuning strategy is used in all experiments in this study.
Figure 8: Origin classification parameter tuning. (a) λ1, λ2, (b) λ, γ

Click here to view


Origin classification

In the origin classification, the number of training and testing samples we selected from D3000 and D4350 is shown in [Table 3]. After cross-validation, the optimal parameters of each algorithm were selected. After ten repetitions of the experiment, we obtained the recall rate. The average classification accuracy and KC are shown in [Table 4] and [Table 5]. It can be seen from the tables that the algorithm adopted in this study has evident advantages, especially for the classification of D3000 data. The classification accuracy of all classes is higher than 90%, and the standard deviation fluctuation is usually small, that is, the model selected in this study is relatively stable and so would be suitable for generalization. This also indicates that under the framework of this algorithm, the spectrum corresponding to the D3000 data of frankincense is more suitable for the origin classification task. [Figure 9] shows the data distribution of the training samples of D3000 and D4350 data in the original space and the distribution in the projection space after feature extraction with the DMLSR algorithm. It is evident that both D3000 and D4350 data have better separability in the projection space, and the intraclass relation is more compact than the original sample and interclass relation more separate. The feature of strong separability can undoubtedly reduce the learning pressure of the subsequent classifier and improve the classification performance. The above experimental results prove that the frankincense data from different regions can be feature extracted by DMLSR to obtain more effective features with less redundant information than the original features, and DCRT can be trained with this low-dimensional feature to obtain a classification model with better generalization ability.
Table 3: Number of training and testing samples for origin classification

Click here to view
Table 4: Accuracy table of origin classification (D3000)

Click here to view
Table 5: Accuracy table of origin classification (D4350)

Click here to view
Figure 9: Visualization of samples from different origins. (a) Visualization of original samples from different origins (D3000). (b) Visualization of projected samples from different origins (D3000). (c) Visualization of original samples from different origins (D4350). (d) Visualization of projected samples from different origins (D4350)

Click here to view


After training and testing, we used the model to predict the origin of frankincense. First, the training model was used to classify the frankincense hyperspectral image pixel by pixel (i.e., label prediction for every pixel except the background), and the origin classification map, as shown in [Figure 10], was obtained, in which three colors represented three origins. There is more than one color in a block of frankincense, because of the misclassification for some pixels in the block. After obtaining the prediction labels of each pixel, the frankincense block was taken as a whole object to make the final prediction. The prediction labels of corresponding pixel points in each frankincense block were counted. The labels with the largest proportion were adopted as the final prediction results of the origin identification of the frankincense blocks. [Figure 11] shows the predicted results of the D3000 data in origin classification. As shown in the figure, each frankincense block obtained a prediction result of origin and a confidence for the corresponding prediction result. The prediction results were all correct. The visual prediction results in [Figure 11] can be applied to the actual frankincense origin classification and identification task.
Figure 10: Origin classification map, D3000 data. (a) India (D3000), (b) Ethiopia (D3000), (c) Somalia (D3000)

Click here to view
Figure 11: Origin predicted results, D3000 data. (a) India (D3000), (b) Ethiopia (D3000), (c) Somalia (D3000)

Click here to view


Level classification

For the classification of level, assuming that the origin of frankincense is known, we selected samples of different levels from the same origin for multiclassification. [Table 6], [Table 7], [Table 8] show the number of selected training and testing samples. Similar to the origin classification, the optimal parameters of each algorithm were selected through cross-validation and the experiment was repeated 10 times. The recall rate, average classification accuracy, and KC were obtained, as shown in [Table 9], [Table 10], [Table 11], [Table 12], [Table 13], [Table 14]. The performance of the algorithm is better than that of SVM and CRT. According to the algorithm flow in this study, the level classification results of the three origins are good on D4350 data, and the classification accuracy of each level is basically above 80%. [Figure 12] shows the visualization of the classification results of samples produced in Ethiopia at different levels in the D4350 data. It can be seen that samples of different levels are closely distributed in the original space. After learning the potential low-dimensional subspace through DMLSR, separability is significantly improved, which is also reflected in the classification accuracy. In addition, owing to the small number of testing samples of each level and the large fluctuation of test precision in repeated experiments, the model is unstable and has a large standard deviation.
Table 6: Number of training and testing samples of different levels of India origin

Click here to view
Table 7: Number of training and testing samples of different levels of Ethiopia origin

Click here to view
Table 8: Number of training and testing samples of different levels of Somalia origin

Click here to view
Table 9: Accuracy table of different levels of India origin (D3000)

Click here to view
Table 10: Accuracy table of different levels of India origin (D4350)

Click here to view
Table 11: Accuracy table of different levels of Ethiopia origin (D3000)

Click here to view
Table 12: Accuracy table of different levels of Ethiopia origin (D4350)

Click here to view
Table 13: Accuracy table of different levels of Somalia origin (D3000)

Click here to view
Table 14: Accuracy table of different levels of Somalia origin (D4350)

Click here to view
Figure 12: Visualization of samples from different levels of Ethiopia origin. (a) Visualization of original samples from different levels (D4350). (b) Visualization of projected samples from different levels (D4350)

Click here to view


After training and testing, we used the model to predict the level of frankincense of different origins. Similar to the origin classification, we used the training model to do the classification pixel by pixel and used the previous strategy to predict the final results of level identification of frankincense. [Figure 13] shows the prediction results based on the D4350 data, which includes level 1 from India (India 1), level 4 from Ethiopia (Ethiopia 4), and level 2 from Somalia (Somalia 2). The red boxes in [Figure 13]a and [Figure 13]c are the frankincense whose level is predicted incorrectly. The visual prediction results in [Figure 13] can also be used in the actual frankincense classification and identification task.
Figure 13: Level predicted results, D4350 data. (a) India (1) (D4350), (b) Ethiopia (4) (D4350), (c) Somalia (2) (D4350)

Click here to view






In this study, DMLSR was selected to extract features from the Frankenstein hyperspectral data. The DMLSR with intraclass compactness graph and manifold regularization can efficiently learn the projective samples with higher separability and less redundant information than the original samples. Then, the classifier using DCRT was adopted for the classification of frankincense's origin and level. DCRT introduces the discriminant regularization term and incorporates SID, which is more sensitive to the spectrum as the measurement method and is more suitable for the frankincense spectral data compared with SVM. Experiments and analysis show that our algorithm framework has excellent classification performance, which is stable in origin classification and has potential for generalization. In addition, the experiments show that in our algorithm framework, different classification tasks need to combine different data sources to achieve better classification and recognition, as the origin classification task uses frankincense's D3000 data, and level classification task uses frankincense's D4350 data.

Financial support and sponsorship

Nil.

Conflicts of interest

There are no conflicts of interest.



 
  References Top

1.
Opinions of the CPC Central Committee and the State Council on promoting the inheritance, innovation and development of traditional Chinese Medicine. Xinhua News Agency; Beijing; 26 October, 2019.  Back to cited text no. 1
    
2.
Huang L. Treasure of traditional Chinese medicine glows with new brilliance, People's Daily. ; Beijing; 13 November, 2019. p. 09.  Back to cited text no. 2
    
3.
Nanjing University of Traditional Chinese Medicine. Dictionary of Traditional Chinese MEDICINE. Shanghai: Shanghai Science and Technology Press; 2006. p. 1991-2.  Back to cited text no. 3
    
4.
Singh GB, Atal CK. Pharmacology of an extract of salai guggal ex-Boswellia serrata, a new non-steroidal anti-inflammatory agent. Agents Actions 1986;18:407-12.  Back to cited text no. 4
    
5.
Xiao J, Liu X, Yan D, Liu R, Zhang W, Tan G. Effect of mastic volatile oil on inhibiting SMMC-7721 cell line proliferation and inducing apoptosis in human liver cancer. Chin J Natl Med 2007;5:68-72.  Back to cited text no. 5
    
6.
Mehrzadi S, Tavakolifar B, Huseini HF, Mosavat SH, Heydari M. The effects of boswellia serrata gum resin on the blood glucose and lipid profile of diabetic patients: A double-blind randomized placebo-controlled clinical trial. J Evid Based Integr Med 2018;23:2515690X18772728.  Back to cited text no. 6
    
7.
Gonzálvez A, Armenta S, Guardia MD. Geographical traceability of “Arròs de Valencia” rice grain based on mineral element composition. Food Chem 2011;126:1254-60.  Back to cited text no. 7
    
8.
Suzuki Y, Chikaraishi Y, Ogawa NO, Ohkouchi N, Korenaga T. Geographical origin of polished rice based on multiple element and stable isotope analyses. Food Chem 2008;109:470-5.  Back to cited text no. 8
    
9.
Zhao H, Guo B, Wei Y, Zhang B. Near infrared reflectance spectroscopy for determination of the geographical origin of wheat. Food Chem 2013;138:1902-7.  Back to cited text no. 9
    
10.
Wang W, Paliwal J. Near-infrared spectroscopy and imaging in food quality and safety. Sensing Instrument Food Quality Safety 2007;1:193-207.  Back to cited text no. 10
    
11.
Liu L, Cozzolino D, Cynkar WU, Gishen M, Colby CB. Geographic classification of spanish and Australian tempranillo red wines by visible and near-infrared spectroscopy combined with multivariate analysis. J Agric Food Chem 2006;54:6754-9.  Back to cited text no. 11
    
12.
Zhang Y, Li W, Li H, Tao R Du Q. Discriminative marginalized least-squares regression for hyperspectral image classification. IEEE Trans Geosci Remote Sens 2019;37:1-14.  Back to cited text no. 12
    
13.
Li W, Tramel EW, Prasad S, Fowler JE. Nearest regularized subspace for hyperspectral classification. IEEE Trans Geosci Remote Sens 2014;52:477-89.  Back to cited text no. 13
    
14.
Li W, Du Q, Xiong M. Kernel collaborative representation with tikhonov regularization for hyperspectral image classification. IEEE Geosci Remote Sens Lett 2015;12:48-52.  Back to cited text no. 14
    


    Figures

  [Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6], [Figure 7], [Figure 8], [Figure 9], [Figure 10], [Figure 11], [Figure 12], [Figure 13]
 
 
    Tables

  [Table 1], [Table 2], [Table 3], [Table 4], [Table 5], [Table 6], [Table 7], [Table 8], [Table 9], [Table 10], [Table 11], [Table 12], [Table 13], [Table 14]



 

Top
 
  Search
 
    Similar in PUBMED
   Search Pubmed for
   Search in Google Scholar for
 Related articles
    Access Statistics
    Email Alert *
    Add to My List *
* Registration required (free)  

 
  In this article
Abstract
Introduction
Research Contents
Research Program
Experimental Des...
Results and Analysis
Conclusion
References
Article Figures
Article Tables

 Article Access Statistics
    Viewed138    
    Printed6    
    Emailed0    
    PDF Downloaded18    
    Comments [Add]    

Recommend this journal