Junhui Li

Near-infrared (NIR) spectroscopy is a promising technique for identifying raw agricultural materials. However, it is rarely used because of its poor discriminant rate. In this study, we took tobacco leaves from five origins as experimental materials. An origin discriminant model by the discriminant partial least squares (DPLS) was established, and the correct discriminant rate of internal cross validation was 76.54%. Origins were divided into three groups. The discriminant model of the three groups was improved, and the correct discriminant rate of internal cross validation was 98.77% along with a 100% external validation rate. We analyzed the characteristics of the three groups’ average spectra under normal variable pretreatment and found they have different absorption characteristics in different regions, and that the classification is information-based. The results show that using information-based classifications can establish a better model, and that main chemical components and NIR spectra can determine whether the classification is information-based, and whether projection based on principal component and Fisher criterion (PPF), can be more effective.

Near-infrared (NIR) light is an electromagnetic wave with a wavelength in the range of 780 to 2526 nm (1). The NIR spectrum absorbs the doubled and combined frequency of hydrogen-containing group vibrations and contains information on the composition of most organic compounds (1). NIR analytical technology is a fast and nondestructive technology for material analysis, and it can conduct two types of analysis: quantitative and qualitative (2). Using NIR qualitative discrimination, the source, type, storage time, grade, and identity of a sample can be determined. The use of NIR spectroscopy to qualitatively discriminate samples has been extensively studied; such studies have considered the varieties of strawberries (3), the origins of jujube (4), the origins of fritillaria (5), the classification of bamboo (6), and explored many other areas (7–11). In addition to NIR analysis technology, methods commonly used to distinguish the characteristics of agricultural products include electronic nose technology (12) and mineral element fingerprint analysis (13), among others.

Qualitative discrimination of agricultural raw materials not only help identify the authenticity, quality, and grade of the agricultural raw materials, but they also facilitate the classification and grading of the market as well as the breeding of varieties, among other processes. Qualitative discrimination of agricultural raw materials is also conducive to stable large-scale industrial production in terms of acquisition and processing. However, in actual production, not many types of at- tributes can be identified (generally only those that differ greatly); the recognition accuracy is not high (for some attribute types, the recognition accuracy is usually less than 90%) (14–17). Although using NIR technology has achieved the aforementioned goals, there are few examples of practical applications.

In this study, we took tobacco leaves from the Sichuan province of China as experimental materials and used discriminant partial least squares (DPLS) to establish a qualitative discriminant model first. However, the correct discriminant rate was poor. We speculated that the classification of origins by administrative divisions is insufficient, or that the classification does not have an information basis. Therefore, we analyzed the similarities between the five origins based on the main chemical components and NIR spectra. Doing this analysis helped establish the discriminant model of the three origin groups. The results showed that using information-based classifications can lead to a better model. This study also analyzed different absorption characteristics in different NIR wavenumber ranges and determined whether the classification is information-based by similarity analysis.

A multipurpose analyzer (MPA)-type Fourier transform NIR spectrometer (Bruker, Germany) was used. The working parameters were as follows: spectrum acquisition range, 12000–4000 cm

The quality characteristics and main chemical components differ greatly among the upper, middle, and lower parts of the same tobacco plant. To eliminate the influence of part differences on the determination of tobacco origins, the middle part of samples collected from different origins in 2014 to 2016 in Sichuan province, China, were selected. Information about the total experimental samples is provided in Table I. Tobacco routine chemical composition data, including the total sugars, reducing sugars, total nitrogen, nicotine, potassium, and chlorine, were obtained by the Sichuan Provincial Tobacco Quality Supervision and Test Station.

Discriminant Partial Least Squares (DPLS)

DPLS originated from partial least squares (PLS) analysis, which is a bilinear model based on an 

 matrix (response variable) (18). This algorithm maximizes the covariance with the 

Although PLS is regarded as a calibration method, it is also used for solving discrimination problems (20). Different from PLS analysis, 

 is a class variable in DPLS. A binary system was used to determine the class to which a sample belongs and a standard is set, where “1” denotes belonging to this class and “0” denotes not belonging to this class (21). When there are three classes or five classes, class variable Y is as shown in Table II. However, the predicted value is often close to 0 or 1 rather than exactly 0 or 1 (22). In general, it can be set to a threshold to discriminate between classes by comparing the relationship between the predicted values and the threshold value (23).

In this study, we use the Caunirs software (China Agricultural University) to establish the origin discriminant models. The correct discriminant rate (CDR) is discriminating one class to within the same class, and the incorrect discriminant rate (IDR) discriminates a class to other classes. The CDR and IDR are defined as follows:

Projection Based on Principal Component and Fisher Criterion (PPF)

The PPF is a projection method that combines principal component analysis with Fisher criterion (24,25). Spectral matrix 

 by a dimension reduction technique called 

 is projected though Fisher criterion. It searches for the projection directions on which the data points of different classes are separated as far as possible, while the data points of the same class are kept close to each other (19).

The PPF method maximizes the between-class distance and minimizes the within-class distance. Not only can the PPF method discriminate between classes, but it can also reflect the similarity between classes (26). The projection of PPF is a circle. In the PPF two-dimensional (2D) projection, the center of each circle represents the projection mean within-class, the within-class projection values reflect the degree of dispersion within the class, and the overlap of the circles reflects the similarity between the classes (19). The distance between the projection graphs can be visually evaluated to ascertain the degree of similarity among the origins; the diameter of the circle can evaluate consistency within an origin (19). This study mainly considers the interclass distances between different origins. In this study, we use the Caunirs software (China Agricultural University) to establish a projection analysis model.

Cluster Analysis and Spectral Preprocessing

Based on the averages of the main chemical components from each origin, SPSS 19.0 software (IBM) was used to perform system cluster analysis. Cluster analysis technology can directly compare the properties of various entities and classify data with similar properties into one class (27). The spectrum is subjected to Savitzky–Golay smoothing preprocessing, first-derivative preprocessing, or standard normal variable (SNV) preprocessing (28). When establishing the DPLS qualitative discriminant models, we used the Savitzky–Golay smoothing preprocessing and first-derivative preprocessing to decrease spectral random deviation, systematic deviation and so on.

When analyzing the absorbance differences in the average spectra among different regions, we used SNV preprocessing to decrease spectral random deviation. SNV pretreatment does not change the original shape of spectrum, and it is convenient for observation of spectral features.

Result of the NIR Qualitative Model for Five Origins

DPLS was used, and the spectral data from 8000 cm

 were selected because of the higher absorbance in this range using the first-derivative and Savitzky–Golay smoothing with 15 points to reduce the spectrum noise (2). Then, based on internal cross-validation (26), the first six principal components were chosen. Table III shows the results of the model.

Table III shows that the total recognition accuracy of the tobacco origins discriminant model is only 76.54%, which does not meet the requirements for practical application. The correct discriminant rates in Panzhihua, Liangshan, Yibin, and Luzhou are low because there are many incorrectly classified samples between Panzhihua and Liangshan as well as between Yibin and Luzhou.

Regarding the reason behind the poor correct discriminant rate, we speculated that the classification of origins by administrative divisions is insufficient or unreasonable or the classification does not have an information basis. Therefore, we studied the similarities between the five origins to obtain a more reasonable or information-based classification based on their main chemical components and spectra to established a better discriminant model.

Similarity Analysis Between Five Origins Based on the Main Chemical Components

The average values of the main chemical components (total sugars, total nitrogen and nicotine) for the five origins were calculated, as shown in Table IV. Using systematic clustering, the contents of total sugars, total nitrogen, and nicotine from the five origins were analyzed. The cluster tree is shown in Figure 1.

FIGURE 1: Illustration of cluster analysis tree.

Table IV shows the following results: Liangshan and Panzhihua tobacco samples have more total sugars, less total nitrogen, and less nicotine. Guangyuan tobacco samples have medium total sugars, more total nitrogen, and more nicotine. Luzhou and Yibin tobacco samples have less total sugars, more total nitrogen and more nicotine. As shown in Figure 1, according to the cluster analysis of the main chemical components, Liangshan and Panzhihua can be in one class, Yibin and Luzhou can be in one class, and Guangyuan can be in a separate class.

Similarity Analysis Between the Five Origins Based on the NIR Spectra

Liangshan, Panzhihua, Guangyuan, Luzhou, and Yibin were labeled with the symbols LSh, PZH, GY, LZ, and YB, respectively. PPF was used, and as in DPLS, the spectral data from 8000 cm

 were selected using first-derivative and Savitzky–Golay smoothing with 15 points. Then, based on the contribution of variance, the first five principal components were chosen. The projection map by PPF is shown in Figure 2.

FIGURE 2: Principal component and Fisher criterion (PPF) projection with 5 principal components.

As shown in Figure 2, the overlap between Liangshan and Panzhihua as well as between Yibin and Luzhou is nearly over half, and the similarity is high. If Liangshan and Panzhihua, Yibin and Luzhou, and Guangyuan can be put into three classifications, the distances between them are all relatively long and the similarity is low.

Result of the NIR Qualitative Model for Three Origin Groups

The results of the similarity analysis between the five origins based on the main chemical components or NIR spectra both show that the similarity between Liangshan and Panzhihua is high, and that the similarity between Luzhou and Yibin is also high. Based on the similarity analysis results, Liangshan and Panzhihua, Luzhou and Yibin, and Guangyuan can be classified as three origin groups, respectively, namely, PL, YL, and GY.

We established a discriminant model for the three origin groups using the same modeling method DPLS and pre-processing method as in the five-origin model. Based on internal cross-validation, the first seven principal components were chosen. Table V shows the qualitative discrimination results of internal cross-validation, and Table VI shows the external validation.

Table V and Table VI shows that, compared with the five origin discriminant model, the discriminant model of internal cross-validation shows a highly correct discriminant rate that increased from 76.54–98.77%, and the correct discriminant rate of external validation reaches 100%.

By comparing the results of the model for five origin groups and the model for three origin groups classified by the similarity of the main chemical components and spectral characteristics, it can be observed that using the information-based classification in distinguishing the origins of tobacco by NIR can establish a better and more reasonable qualitative discriminant model. This conclusion may be extended to other agricultural raw materials.

As shown in Figure 3, Figure 4, Table VII, and Table VIII, there are significant differences in the average spectra among Liangshan and Panzhihua, Luzhou and Yibin, and Guangyuan. In 4250–4350 cm

, the spectra mainly contain the absorption of C-H groups typically on cellulose, and the PL has a lower absorbance, while the YL and GY have higher absorbance. In 4700–4800 cm

, the spectra mainly contain the absorption of O-H groups typically on starch and other carbohydrates, and the PL has higher absorbance, while the YL and GY have lower absorbance.

, the spectra mainly contain the absorption of N-H groups typically on protein. The PL has a lower absorbance, the YL has a middle absorbance, and the GY has a higher absorbance.

Based on the information regarding the composition of organic groups and related substances absorbed by NIR light in different frequency ranges and the absorption characteristics of different origin groups of tobacco in different bands, it can be inferred that the PL tobacco samples may have less cellulose, more starch and other carbohydrates, and less protein; the YL tobacco samples may have more cellulose, less starch and other carbohydrates, and a middle amount of protein; and the GY tobacco samples may have less cellulose, less starch and other carbohydrates, and more protein.

The chemical composition characteristics of different origin groups are obtained by spectrum absorption characteristic analysis, and they are consistent with the characteristics of routine chemical composition analysis obtained by chemical measurement, such as carbohydrates. Average spectral information further verifies that the division into three origin groups is an information-based classification; it can also be used for determining whether the classification has an information basis. However, methods such as similarity analysis may be more effective because of the serious overlap of NIR spectra. At the same time, we can also use NIR spectroscopy to obtain content information of the major chemical groups in agricultural raw materials.

This study established a qualitative model for five origin groups, and the correct discriminant rate was poor. Based on the similarity of the main chemical components and near-infrared spectra of the different origin groups, a discriminant model for three origin groups was established, and the correct discriminant rates of internal cross-validation and external validation are both greatly improved.

The analysis of absorption characteristics in the spectra showed that they can be used for determining whether the classification has an information basis by analyzing the preprocessing average spectra. At the same time, we can also use near-infrared spectroscopy to obtain content information of the major chemical groups in agricultural raw materials.

Using the information-based classification can lead to a better model, and the classification algorithm is not most important. The chemical composition or spectral characteristics can be used to determine whether the classification has an information basis. Compared to chemical composition analysis, using NIR spectroscopy is more convenient, and spectrum similarity analysis methods, such as PPF, can be more effective because of the serious overlap of near-infrared spectra. The proposed method has possible applications for other agricultural products.

This research is supported by the China National Key Research and Development Program Subtopic (Project No.: 2016YFD0700304) and the Major Science and Technology Project in Sichuan Province, China (Project No.: SCYC201810).

Basic and Application of Near Infrared Spectroscopy

 (China Light Industry Press, Beijing, China, 2005), pp. 8–39.

Principle, Technology and Application of Near Infrared Spectroscopy

 (China Light Industry Press, Beijing, China, 2013), pp. 120–144.

(3) X.Y. Niu, L.M. Shao, Z.L. Zhao, X.Y. Zhang, 

(4) W.J. Wang, X.G. He, X.T. Yang, S.L. Wang, J.Y. Wang, 

(5) Y. Meng, S.S. Wang, R. Cai, B.H. Jiang, and W.J. Zhao, 

(6) Z. Yang, K. Li, M.M. Zhang, D.L. Xin, and J.H. Zhang, 

(7) H. Li, F.R. van de Voort, A.A. Ismail, J. Sedman, R. Cox, C. Simard, and H. Buijs, 

(8) J.W. Zhao, Q.S. Chen, H.D. Zhang, and M.H. Liu, 

(9) C.Y. Wang, B.R. Xiang, and W. Zhang, 

(10) S.M. Tan, R.M. Luo, Y.P. Zhou, H. Xu, D.D. Song, T. Ze, T.M. Yang, and Y. Nie, 

(11) D.M. Musingarabwi, H.H. Nieuwoudt, P.R. Young, E. Bickong, A. Hans, and M.A. Vivier, 

(12) B. Shi, L. Zhao, R.C. Zhi, X.J. Xi, D.Z. Zhu, 

(14) X.L. Li, Y.M. Tang, Y. He, and X.F. Ying, 

(15) F. Cao, D. Wu, Y. He, and Y.D. Bao, 

(16) H.R. Wang, X.L. Chen, W.J. Li, and J.L. Lai, 

(17) D. Eisenstecken, A. Panarese, P. Robatscher, C.W. Huck, A. Zanella, and M. Oberhuber, 

(19) L.L. Luan , Y.H.Wang , X.Y. Li, W.Y. Hu, K. Li, J.H. Li, K. Yang, R.X. Shu, L.L. Zhao, and C.L. Lao, 

(21) H. Qin, H.R. Wang, W.J. Li, and X.X. Jin, 

(22) S.W. Lindstrom, P. Geladi, O. Jonsson, and F. Pettersson, 

(23) M.R. Almeida, D.N. Correa, W.F.C. Rocha, F.G.O. Scaf, and R.J. Poppi, 

Principal Component Analysis, 2nd Ed., Springer Series in Statistics

Pattern Classification and Scene Analysis

(26) L. Zhang, X.H. Tang, X. Ma, Y.Y. Qian, L.P. Wang, Y.D. Wen, Y. Wang, H.H. Zhang, L.L. Zhao, and J.H. Li, 

Application of Modern Near-Infrared Spectroscopy Analysis Technology in Drug and Food Quality Evaluation System

 (Higher Education Press, Beijing, China, 2009), pp. 74–79.

(29) Y.L. Yan, L.L. Zhao, D.H. Han, and S.M. Yang, 

Basics and Applications of Near Infrared Spectroscopy

 (China Light Industry Press, Beijing, China, 2005).

 are with the College of Information and Electrical Engineering at China Agricultural University in Beijing, China. 

 is with the Sichuan Tobacco Company of CNTC in Chengdu, China. 

 is with the Customs Technology Center of Chengdu in Chengdu, China. 

 is with the Panzhihua Tobacco Company of CNTC in Panzhihua, China. Direct correspondence to Junhui Li at 

Articles

Traditional qualitative analysis of agricultural materials using near-infrared spectroscopy can be improved using information-based classification methods, such as projection based on principal components and the Fisher criterion (PPF).

Using Information-Based Classifications to Distinguish Characteristics of Raw Agricultural Materials by Near-Infrared Spectroscopy

The establishment of quantitative models based on the near-infrared (NIR) spectroscopic analysis of plant samples plays an important role in improving both the scope of the models and the accuracy of prediction. In this study, four types of tobacco samples and their mixed products were selected as research objects. Based on the modeling methods of the different types of tobacco, this technique could provide a new method for tobacco quality management as well as provide a new discriminant method for other agricultural products.

	To improve the scope of models and the accuracy of their prediction, the establishment of quantitative models for different types of plants plays an important role. The quality characteristics of common food and cash crops, including wheat, rice, corn, tea, and tobacco, among others, differ because of genetic differences. Among these crops, tobacco exhibits typical characteristics of plants and can be divided into several types, including flue-cured tobacco, burley tobacco, dark sun-cured tobacco, and oriental and aromatic tobacco. Moreover, total nitrogen and total alkaloid contents are two important indicators of quality evaluation in tobacco and its products. The total nitrogen content is determined by the classical Kjeldahl method, and the total nicotine content is determined by continuous flow analysis (1,2). However, both methods are limited by the need for complex preprocessing, and the testing processes are time-consuming and can delay results; therefore, these methods do not meet the rapid detection and quality control needs of scientific research or the on-site quality acceptance of tobacco leaves. At present, local regression modeling and consensus modeling are used to analyze homogeneous plant samples that exhibit regional and temporal differences (3–7). Beyond ensuring the applicability of models, significantly improving their accuracy is difficult, so it is important to develop a fast and accurate quantitative analysis method for improving the applicability of the models.

	As a nondestructive and rapid analytical method, near-infrared (NIR) spectroscopy has been widely used for the qualitative analysis of various organic feedstocks and for physical and chemical quantitative analyses across the food, pharmaceutical, and agricultural industries (8,9).

	A variety of chemical composition tests using quantitative NIR analysis have recently been carried out on flue-cured tobacco, burley tobacco, oriental and aromatic tobacco, and other types of tobacco (10–17). Previously, Tonini reported an NIR model of the total sugar content in both flue-cured tobacco and oriental and aromatic tobacco (18). The results were as follows: For flue-cured tobacco, the standard error of calibration (SEC) was 0.802 and the standard error of prediction (SEP) was 1.208; whereas for oriental and aromatic tobacco the SEC was 0.883, and the SEP was 0.976. These results showed that the single models were more accurate than were the hybrid models. Then, new pattern recognition methods by Zhang (19) using a partial least squares support vector machine (PLS-SVM) approach and by Shao (20) using wavelet transform combined with an artificial neural network (WT-ANN) were used to establish the quantitative and discrimination models and showed an excellent prediction performance for rapid and accurate analysis of routine chemical compositions in tobacco. To date, however, most studies essentially report quantitative analysis methods of the main chemical constituents of flue-cured tobacco. The comparison of NIR single and hybrid models for total nitrogen and total alkaloid contents in other tobacco types has not been reported, which strongly limits the applicability of quantitative models.

	In this study, different types of tobacco were used as research samples. The total nitrogen content was determined by the Kjeldahl method, and the total alkaloid content was determined by continuous flow analysis. The spectra were determined by Fourier transform NIR spectroscopy and were preprocessed by the first derivative and smoothing methods. Single and hybrid quantitative analysis models for total nitrogen and total alkaloid contents in flue-cured tobacco, burley tobacco, dark sun-cured tobacco, oriental and aromatic tobacco, and mixed tobacco were ultimately established successively through the use of PLS and PLS-discriminant analysis (PLS-DA).

	All the experimental samples of different tobacco types were collected at the Beijing Third Class Tobacco Supervision Station within a three-year period and are shown in Table I.

	A multipurpose analyzer with a near-infrared integrating sphere diffuse reflection accessory in reflectance mode (Bruker Optics Inc.) was used to measure spectra. In the experiments, wavenumbers ranging from 12,000 to 3500 cm

 were measured at a digitization interval of approximately 8 cm

. The scan rate was 64 within 30 s, and all measurements were performed at room temperature. OPUS 7.0 quantitative analysis software (Bruker Optics Inc.) was used. The average of three normal spectra per sample was used.

	The total nitrogen in the tobacco samples was determined using a B-339 nitrogen analyzer (Buchi Labortechnik AG) in accordance with the Chinese tobacco industry YC/T 33-1996 standard. The total alkaloid content in the tobacco samples was determined using an AA3 continuous flow analyzer (Bran + Luebbe) in accordance with the Chinese tobacco industry YC/T 160-2002 standard. These data were used for subsequent modeling.

	PLS was used to establish quantitative models of total nitrogen and total alkaloids for flue-cured tobacco, burley tobacco, dark sun-cured tobacco, oriental and aromatic tobacco, and mixed tobacco (21). Internal cross validation was used to determine the statistical parameters of the predictive NIR spectroscopy models. The calibration models were validated by verification sets. The SEP and mean of the SEP between the measured values and the predicted values of total nitrogen and total alkaloids were used as the evaluation parameters (22,23).

	PLS-DA is a supervised pattern recognition method based on PLS and a combination of PLS regression and the discrimination analysis for classification tasks (24,25). The method uses a priori classification knowledge to establish a classification model using PLS-DA and then uses that model to discriminate the attributions of the samples to be tested.

Establishment of Single and Hybrid Models

	For each type of tobacco, 150 samples were collected. Based on the distribution of total nitrogen and total alkaloid concentrations, 120 samples were selected for the modeling set, and the remaining samples (comprising 30 samples of mixed tobacco) were used for the validation set. Summaries of the statistical analysis results of the partially optimized PLS of the models of the different tobacco types for the total nitrogen and total alkaloid contents are shown in Tables II and III.

 of the quantitative NIR models of total nitrogen and total alkaloids among flue-cured tobacco, burley tobacco, dark sun-cured tobacco, oriental and aromatic tobacco, and multitype tobacco were greater than 0.90; the SEC values of the calibration sets were lower than 1; and the calibration models had better predictive ability.

	The external prediction of the single and hybrid quantitative calibration models for total nitrogen and total alkaloids was carried out by using the 30 validated samples. The predictive results of the different quantitative NIR analysis models for total nitrogen and total alkaloids are shown in Tables IV and V; the results are reported as SEP values. The prediction results of total nitrogen from the different models of tobacco are shown in Tables VI and VII; the results are reported as the means of the SEP. Among these models, S-K, S-Mu, H-S, S-Mi, and H-Mi represent single models predicting known samples, single models predicting multitype samples, hybrid models predicting known samples, single models predicting mixed samples, and hybrid models predicting mixed samples, respectively.

	As shown in Tables IV and VI, in the forecast of the quantitative model of the total nitrogen content in the different types of tobacco (the data in the table are bold and italicized), the single models can predict known samples of total nitrogen. The SEP values of flue-cured tobacco, burley tobacco, dark sun-cured tobacco, and oriental and aromatic tobacco were lower than those of the single models predicting multitype samples, demonstrating that the S-K models are better than the S-Mu models. Further, the error rate for predicting total nitrogen was reduced by 42.5% and for total alkaloids by 63.8%, indicating that single models could predict only known samples but not mixed samples.

	The comparison of the S-Mi and H-Mi models shows that the mean of the SEP was 0.210 for S-Mi, which was greater than that for H-Mi (0.185), so the accuracy of the single models predicting mixed samples was clearly weaker than that of the hybrid models. Compared with that of H-K, the error rate of H-Mi was reduced by 11.9% for total nitrogen and by 69.5% for total alkaloids. This finding indicates that the hybrid models are more suitable and broadly applicable; they can be used for the quantitative analysis of mixed products and are suitable for the quantitative analysis of other agricultural plants (for example, different varieties of corn and different modulation methods of tea). In particular, regarding the comparisons of H-K and S-K with S-Mu, for unknown types of tobacco, the H-K model can be applied, but its predictive effect is weaker than that of the S-K model but stronger than that the S-Mu model.


	Figure 1: Comparison of single models predicting known samples and hybrid models predicting known samples of the total nitrogen of flue-cured tobacco and burley tobacco: (a) flue-cured tobacco and flue-cured tobacco, (b) burley tobacco and burley tobacco, (c) multitype tobacco and flue-cured tobacco, and (d) multitype tobacco and burley tobacco.

	To verify the predictive effects, we further took flue-cured tobacco and burley tobacco as examples in this work and comparatively analyzed the total nitrogen and total alkaloids from the single model prediction of flue-cured tobacco (or burley tobacco) and hybrid model prediction of flue-cured tobacco (or burley tobacco) as well as from the single model prediction of mixed tobacco and hybrid model prediction of mixed tobacco. The correlations between the measured values and the predicted values of the verification sets are shown in Figures 1–4.


	Figure 2: Comparison of single models predicting mixed samples and hybrid models predicting mixed samples of the total nitrogen of flue-cured tobacco and burley tobacco: (a) flue-cured tobacco and mixed tobacco, (b) burley tobacco and mixed tobacco, and (c) multitype tobacco and mixed tobacco.

	As shown in Figures 1–4, the total nitrogen and total alkaloids of the different tobacco types were concentrated on both sides of the line, and a high degree of analysis accuracy occurred with respect to the relationship between the measured values and the predicted values of the verification sets. Figures 1 and 3 show that the predictive effect of the single models is clearly better than that of the hybrid models. By contrast, Figures 2 and 4 show that the hybrid models constructed in this study have a wider predictive range for predicting mixed samples.


	Figure 3: Comparison of single models predicting known samples and hybrid models predicting known samples of the total alkaloids of flue-cured tobacco and burley tobacco: (a) flue-cured tobacco and flue-cured tobacco, (b) burley tobacco and burley tobacco, (c) multitype tobacco and flue-cured tobacco, and (d) multitype tobacco and burley tobacco.

Model Discriminant Analysis of Qualitative and Quantitative Combinations of Different Types of Tobacco

	In practical applications, unknown sample types are often present. In general, one way of estimating sample types involves the use of hybrid models, and then, based on the predicted results and the empirically determined total nitrogen and total nicotine contents of the tobacco, samples suitable for classification are selected. To improve the accuracy of quantitative NIR analysis, the use of qualitative discrimination followed by a single quantitative analysis model can be employed.


	Figure 4: Comparison of single models predicting mixed samples and mixed models predicting mixed samples of the total alkaloids of flue-cured tobacco and burley tobacco: (a) flue-cured tobacco and mixed tobacco, (b) burley tobacco and mixed tobacco, and (c) multitype tobacco and mixed tobacco

	PLS-DA, which is based on both qualitative and quantitative algorithms, can be used to determine the chemical composition of unknown samples. The PLS-DA was performed before the second derivative and smoothing (17) transformations, and cross validation was used to determine that the value of the factor was 10. The qualitative discriminant analyses between the different types of tobacco are shown in Figure 5, and the discriminant rates are shown in Table VIII.


	Figure 5: Prediction results based on PLS-DA of the various tobacco types: (a) flue-cured tobacco, (b) burley tobacco, (c) dark sun-cured tobacco, and (d) oriental tobacco.

	Figure 5 shows that the qualitative identifications of the flue-cured tobacco were all correct, as were those of other sample types. Two of the burley tobacco types were incorrectly identified, but the others were correctly identified; two of the other sample types were mistaken as burley tobacco. The qualitative identifications for the dark sun-cured tobacco were all accurate, but the other sample types of two burley samples were mistaken as dark sun-cured tobacco. The qualitative identifications for the oriental tobacco were also all correct; among the other sample types, one dark sun-cured tobacco sample was mistaken as oriental and aromatic tobacco, but the rest were correctly identified.

	Table VIII shows that the discriminant rates of the different tobacco types were all above 95%, effectively improving the prediction accuracy of the quantitative NIR analysis. These results also confirm that significantly improving the accuracy of plant regression and quantitative NIR analysis methods is difficult.

	This article describes the use of PLS and PLS-DA to establish quantitative single and hybrid NIR spectroscopy models for estimating total nitrogen and total alkaloid contents. The means of the SEP from the S-K, S-Mu, S-Mi, H-K, and H-Mi models were 0.123, 0.214, 0.140, 0.210, and 0.185 for total nitrogen and 0.192, 0.530, 0.236, 0.463, and 0.141 for total alkaloids; the error rates were clearly reduced. The single models were more accurate and exhibited stronger predictive ability; thus, these models can be used to predict known sample types. The hybrid models were applied to mixed tobacco. For unknown types of tobacco, the use of qualitative discrimination followed by the S-K model for unknown sample types effectively improved the prediction accuracy of the quantitative NIR models. This technique could constitute a new method for tobacco quality management. The quantitative models are also applicable to the NIR spectroscopic analysis of other plant samples and represent a new discriminant method for other agricultural products whose appearance is unclear or misleading for various reasons.

 (Tsinghua University Press, Beijing, China, 1994).

(Chemical Industry Press, Beijing, China, 2004).

	(3) T. Naes, T. Isaksson, and B. Kowalski, 

	(4) Z. Wang, T. Isaksson, and B.R. Kowalski, 

	(8) C.C. Fagan, C.D. Everard, and K. McDonnell, 

	(10) W. McClure, K. Norris, and W. Weeks, 

	(15) J.P. Wold, D. Airado-Rodriguez, and A.K. HoltekjØlen, 

Spectrochim. Acta A Mol. Biomol. Spectrosc. 

	(17) J.P. Zhang, W.Y. Xie, and R.X. Shu, 

	(18) A. Tonini, S. Commella, and G. Mellone, 

	(21) S. Wold, M. Sjöström, and L. Eriksson, 

Technology and Application of NIR Spectra Analysis

 (China Light Industry Press, Beijing, China, 2013).

Qualitative Analysis Methods of Near Infrared Spectroscopy

 (Chinese Light Industry Press, Beijing, 2013).

 are with the College of Information and Electrical Engineering at China Agricultural University in Beijing. 

 are with the Beijing Third Class Tobacco Supervision Station in China. Direct correspondence about this article to Yuqing Yang at 

The establishment of quantitative models based on the near-infrared (NIR) spectroscopic analysis of plant samples plays an important role in improving both the scope of the models and the accuracy of prediction. This technique could provide a new method for tobacco
quality management and provide a new discriminant method for other agricultural products.

Junhui Li

Articles

Using Information-Based Classifications to Distinguish Characteristics of Raw Agricultural Materials by Near-Infrared Spectroscopy

Exploring the Applicability of Quantitative Models Based on NIR Reflectance Spectroscopy of Plant Samples