CHINESE JOURNAL OF ANALYTICAL CHEMISTRY Algorithms, Strategies and Application Progress of Spectral Searching Methods

CHINESE JOURNAL OF ANALYTICAL CHEMISTRY Volume 42, Issue 9, September 2014 Online English edition of the Chinese language journal Cite this article as: Chin J Anal Chem, 2014, 42(9), 1379–1386. REVIEW Algorithms, Strategies and Application Progress of Spectral Searching Methods CHU Xiao-Li*, LI Jing-Yan, CHEN Pu, XU Yu-Peng Research Institute of Petroleum Processing, Beijing 100083, China Abstract: In recent years, many modern spectral databases for complex materials (such as soil, feed, forensic evidence materials, pharmaceuticals, oils, and so on) have been established on the basis of molecular spectroscopy (UV, infrared, near infrared, Raman and fluorescence), which are playing more and more important roles in the agricultural, industrial production and science research. Spectral searching method is one of the key techniques to make full use of the molecular spectral database. This paper reviewed the progress in the basic and modified algorithm, strategy and application of molecular spectral searching methods, and discussed the scientific and technological problems that need attention and further research. Key Words: Molecular spectroscopy; Spectral searching algorithm; Correlation coefficient; Moving window; Review 1 Introduction In recent years, with the improvement of equipment manufacturing and the popularization of chemometric methods, molecular spectroscopy analysis technology, especially the infrared, near infrared and Raman spectroscopy, has been widely applied in many fields due to the advantages of convenient test, fast speed, rich information, on line analysis and so on. By the methods of pattern recognition, clustering or recognition of molecular spectra was used for the analysis of complex system, such as oil, grain, fruit, and medicines[1]. In chemometrics, as shown in Fig.1, the modern pattern recognition methods of molecular spectroscopy analysis include three categories[2]: (1) unsupervised methods, such as principal component analysis, clustering method, K-means clustering method, and self-organizing neural network; (2) supervised methods, such as Linear discriminat analysis (LDA), Soft independent modeling of class analogy (SIMCA), Discriminant partial least squares (DPLS), and Support vector machine (SVM); all the methods above are based on the sample types, when a new sample is added to the database, the qualitative models must recalibrated. (3) spectral searching methods, such as correlation coefficient, cosine, Euclidean distance, and spectral information divergence. Based on the spectra of unknown samples, spectral searching methods can qualitatively and quantitatively analyze samples by searching the most similar one or more samples from the built spectral library. Previously, the spectral searching methods were mainly used for spectral identification of pure compounds, such as Sadtler and Aldrich infrared spectrum database. In recent years, the modern molecular spectra databases of complex mixtures established gradually in many fields (such as soil, feed, evidence materials, medicines, and oil)[3‒5], and spectral searching algorithm was more and more popular[6]. New searching algorithm and strategies emerged, and the accuracy and reliability of the spectral searching results were significantly improved. Compared with the unsupervised and supervised pattern recognition method, the spectral searching method has the advantages of simple operation, visual information and convenient maintenance of library, which plays an important role in practical applications. In this review, novel molecule spectral searching algorithms, strategies and their applications are introduced. The scientific and technological challenges are also discussed. Received 13 May 2013; accepted 2 July 2014 * Corresponding author. Email: cxyuli@sina.com Copyright © 2014, Changchun Institute of Applied Chemistry, Chinese Academy of Sciences. Published by Elsevier Limited. All rights reserved. DOI: 10.1016/S1872-2040(14)60768-4 CHU Xiao-Li et al. / Chinese Journal of Analytical Chemistry, 2014, 42(9): 1379–1386 Fig.1 Classification diagram of pattern recognition methods 2 Basic spectral searching algorithm For the x spectrum of unknown sample, the aim of spectral searching is to find the most similar one (or more) sample spectrum to x, based on certain algorithms and rules. If the properties matrix Y is known in spectral library, the properties of unknown sample can be predicted according to the spectral searching results. In details, x represents unknown spectrum, organized as 1×m vector, m represents wavelength points; R represent all spectra in the library, organized as n × m matrix, where n is the number of sample; rj represents the jth sample in the spectral library, organized as 1×m vector, j = 1, 2, ..., n; Y represents properties of corresponding spectra in library, organized as n × p matrix, p represents the number of properties; yj represents property value of the jth sample in spectral library, organized as 1 × p vector. In order to obtain the optimal results, spectral pretreatment and wavelength range selection are needed. Pretreatment methods include derivative, vector normalization, standardization, and wavelet transform. Based on chemistry knowledge and mathematics, wavelength selection methods can identify those spectral intervals which are strong characteristic, high in signal to noise ratio, and less vulnerable to external factors. There are several references related to the common used spectral pretreatment and wavelength selection methods[7]. 2.1 Distance based algorithm The basic principle of this algorithm is that the more similar the spectra of two samples are, the shorter the distance between two samples in the spectra. There are various forms of spectral distance, and the absolute distance is the Absolute distance between the sample spectrum x and the jth sample, represented as rj in the spectra library, can be expressed as following: d(x, rj) = ∑|x – rj| (1) The Euclidean distance, known as the least square distance, is defined by the formula: d(�,� )= 2.2 �-� T �-� Similarity algorithm (2) There are two parameters for evaluation of the similarity between two spectra: cosine and correlation coefficient. The cosine between x and rj is expressed as follows: �� T cos �,� = ��T � � T (3) The basic principle of this algorithm is that the smaller the cosine angle is, the greater the similarity of two samples is. If the two spectra are entirely identical, the cos(x, rj) = 1, the two samples in the pattern space get close to one point; if the two spectra are completely different, cos(x, rj) = 0. The correlation coefficient between the x and rj is expressed as follows: (� − �)T (� − � )T �,� = (� − �)(� − �)T (� − � )(� − � )T (4) The x and rj are average values of x and rj , respectively. If the value is closer to 1, it means that the two spectra are more similar; if the value is closer to 0, it means that the two spectra are more different. 2.3 Algorithm based on information theory Spectral information divergence (SID) [8] can be used to evaluate spectral similarity by relative entropy of spectral information: SID(x, rj) = D(x||rj) + D(rj||x) (5) where, D(x||rj) is the relative entropy with rj to x, and D(rj||x) is the relative entropy with x to rj: D �||� = D � ||� = � lg ( � =1 ) , =1 , lg (6) , (7) CHU Xiao-Li et al. / Chinese Journal of Analytical Chemistry, 2014, 42(9): 1379–1386 where, are the probability vector of spectrum x and rj, q = x/(∑mi = 1 xi), pj = rj/(∑mi = 1 xj,i), respectively. 2.4 Extent similarity algorithm Extent similarity algorithm is simplified based on the evolution of similarity system theory[9,10]. Parameter Q of extent similarity reflects the degree of similarity, which represents the average relative difference between two spectra: � min⁡ (� , , ) 1 =1− (1 − ) max⁡ (� , , ) � =1 (8) The algorithm compares the spectral intensity at each wavelength point; therefore, it is sensitive to the relative difference in wavelengths. If the Q value gets closer to 1, it shows that the two spectra are more similar; if the Q the value gets closer to 0, it shows that the two spectra are more different. 2.5 Jaccard similarity algorithm Jaccard similarity is a matching algorithm based on spectral peaks using binarization processing [11,12], which calculates the characteristic peaks proportion of intersection as a share of union: � � �= = + + � � (9) where, p is the number of corresponding wavelength points of 1 after the binarization of x and rj; q is the number of corresponding wavelength points of 1 after the binarization of x plus the number of corresponding wavelength points of 0 after the binarization of rj; r is the number of corresponding wavelength points of 0 after the binarization of x plus the number of corresponding wavelength points of 1 after the binarization of rj. 2.6 Improvement and application of spectral searching algorithm Based on the absolute distance (spectral difference), and considering the effects of eliminating absorption intensity on the difference, Meng et al. established a new method to calculate the similarity S of ultraviolet spectra: � � �= = + + � � (10) Compared with the cosine method and correlation coefficient, this method is highly sensitive to spectral differences, and can overcome the disadvantages of broadband UV absorption spectrum to a certain extent. The similarities and differences between the traditional Chinese medicines, were rapidly and accurately identified with this method, by which the production process of traditional Chinese medicine injections was also monitored online[13]. Li and Tang improved the method of the similarity S, by increasing the weights of the calculation. The sensitivity of S was increased by changing key spectral wavelength range. The method was used to identify the similarity of near infrared spectra or examine anomalies stability on UV spectra of Danshen injection[14,15]. Khan[16] proposed a new Raman spectroscopy specific spectral similarity metric, called spectral linear kernel, which performed better in comparison with standard spectral searching methods. The recognition result was better than the traditional Euclidean distance and cosine. Considering the spectral repeatability of sample during the testing process, Plugge et al[17] put forward the consistency test (Conformity Index, CI) method based on absolute distance, and its essence was a weighted absolute distance method. It uses the average spectra of a group of repetitive spectrum instead of library spectra rj, and each weight of wavelength point is the reciprocal of the repetitive spectral standard deviation σj: � − , CI = MAX( ) �, (11) CI means spectral reproducibility, which is usually 3 to 5 times the standard deviation. The method was used to detect changes in physicochemical properties of ampicillin trihydrate, which could control the production process and guarantee the consistency of product quality. Ritchie investigated the accuracy, precision, robustness and consistency of CI method. The results showed that the method met the current desired requirement; it could be accepted by the modern strict guidelines[18]. Feng et al[19] took advantage of the consistency test method to measure the authenticity of drug quality by near infrared (NIR) spectroscopy, and the NIR database for consistency test was applied for hundreds and thousands of drugs[20,21]. For repetitive spectra or multiple samples of one class in library, Thermo used principal component analysis (PCA) to calculate the difference e between sample and library spectra. A modified Euclidean distance similarity matching value (Similarity match value, SMV) was defined as follows: |�| SVM = (1 − ) × 100 |�| (12) For near infrared spectra of complex mixtures, if the main component in the mixture is the same, it is difficult to identify the difference between the samples through the traditional spectral searching methods. Nie et al. combined NIR spectroscopy with the SMV method to identify differences quickly and conveniently in Wujibaifeng pills from Tongren and other manufacturers[22]. Tao et al[23] combined near infrared diffuse reflectance spectra with the SMV method to characterize the stability of cigarette quality, which could perform quick and large samples detection in the preparation of tobacco, providing a new technique for cigarette processing quality control. A textile fibre infrared spectral library containing 1000 samples was established by Lu et al[24] using CHU Xiao-Li et al. / Chinese Journal of Analytical Chemistry, 2014, 42(9): 1379–1386 attenuated total reflectance (ATR) measurement, and rapid detection of fiber types was realized on the basis of the spectral searching function. To detect 18 kinds of common plastic resin, Wang et al[25] established a standard infrared spectral library containing 513 samples, by which the plastics were rapidly identified. The correlation coefficient and the cosine were the most commonly used library searching methods[26,27]. For example, based on 940 paint infrared spectra obtained from 287 car body paint samples, Chen et al. established a vehicle body paint infrared spectra comparison database by the correlation coefficient method. This method was used in the traffic accident scene investigation for rapid determining the escaping vehicles[28]. He et al[29] established a Drug terahertz spectral database containing 38 samples, which was expected as a complementary means for the drug detection. Guedes et al[30] used correlation coefficient method to establish a micro Raman spectroscopy database for identification of airborne pollen types. In addition, correlation coefficient and cosine were most common methods in hyper spectral remote sensing image database object identification[31]. The traditional correlation coefficient and cosine emphasize the integral similarity of spectrum. To figure out the emphasis on the spectral detail difference, several improved methods for correlation coefficient and cosine were applied. Two-step cascade library was another common strategy, proposed by Blanco et al[32] to improve the searching accuracy, which was successfully used for the identification of NIR spectra of medicinal materials. Computing correlation coefficient by feature selection of intervals is also an extremely effective method. Wang used spectral interval correlation coefficient method to fast discriminate whether sildenafil citrate was added into the traditional Chinese medicine capsule, with an overall screening accuracy rate at around 95%[33,34]. Xu[35] divided the whole spectrum into several regions and calculated the correlation coefficient or cosine of each region respectively. This correlation coefficient method, called the correlation coefficient array, increased the difference between spectra to a certain extent. The authenticity test of traditional Chinese medicines showed that the method was superior to the traditional correlation coefficient. Griffiths et al[36] proposed a weighted spectrum correlation coefficient matching algorithm, which could effectively overcome the interference signals existing in the infrared spectra. Reasonable recognition results were obtained in the application of Open-path infrared spectroscopy to monitor air pollutants. Based on the concept of moving window, Chu et al[37] developed a new method, named moving window correlation coefficient method. This approach selects a spectral window that starts at the kth spectral channel and ends at the (k + w − 1)th spectral channel (the window width is w), by moving the spectral window successively along the equally spectral data to construct a series of moving window (a total of n – w + 1 windows), and then calculates corresponding correlation coefficients in each moving window using the traditional correlation coefficient formula. For this method, the window width is a very critical parameter. A narrow moving window would be useful to distinguish tiny differences between two spectra, but the risk is that the repetitive spectra of the same sample collected in different dates may not be accurately identified because of spectral measurement errors and instrumental variations. A wide moving window could reduce the impact of external testing influence such as temperature and humidity, but two different samples with very subtle differences may not be effectively distinguished. Thus, the window width needs optimizing in the practical applications according to the differences of samples and the spectra measurement conditions. Chu and Li established crude oil near infrared and infrared spectral identification database by moving window correlation coefficient method, which could accurately identify the type of crude oil[38,39]. Li also used moving window correlation coefficient to identify crude oil 2D IR spectra, and the matrix window correlation coefficient method could accurately identify the mixed crude oil of low proportion[40]. Guo et al[41] made use of near infrared spectroscopy and moving window correlation coefficient method to determine the end point of traditional Chinese medicine extraction process. Compared with the original moving standard deviation method, this method could weaken the effect of baseline drift to a great extent, showing better robustness of testing. Ramirez-Lopez et al[42] improved the traditional difference spectra by presenting a surface difference spectrum (SDS). The results indicated that the SDS approach was a suitable method for computing distances in the spectral space, which could be used in proximal soil vis-NIR sensing applications. 3 Spectral retrieval strategy and application In order to obtain accurate and fast spectral searching results, several new and improved search strategies on the above basic and improved searching algorithms were presented for specific applications. Only using one library searching method sometimes may result in less robust recognition. To solve this problem, ensemble or consensus strategy is applied. The basic ideas of these strategies are to establish rules respectively by variety of searching algorithms, and then these rules are used to recognize the sample synchronously, and the recognition result comes from the final hit rate or weighted value. The search strategy reduces the dependence of search results on a particular algorithm, thus the stability of search results is improved. Himmelsbach et al[43] established an ATR infrared spectral database to identify the extraneous matter in the cotton. The CHU Xiao-Li et al. / Chinese Journal of Analytical Chemistry, 2014, 42(9): 1379–1386 database contained 601 sample spectra, including contaminants typically classified as “trash”, cotton plant parts (hull, shale, seed-coat fragments, bract, cacyx, leaf, bark, sticks, and stems), and grass plant parts (leaf and stem); synthetic materials (plastic bags, film, rubber, bale wrapping and strapping); organic materials (other fibers, yarns, paper, feathers, and leather). However, the recognition accuracy went down significantly when the database used in recognizing spectra of cotton from different regions, picking period or spectra collected with different conditions. Loudermilk et al[44] used the consensus strategy to integrate results of 6 kinds of common library spectra searching methods with satisfactory result. The Euclidean clustering, correlation coefficient and spectral information divergence were integrated by Kong et al[45] to apply in Airborne Hyper spectral Remote Sensing Image of China's practical modular imaging spectrometer system (OMIS) and mineral spectral library of American Geological Survey (USGS). The results showed that it had stronger spectral discrimination and smaller uncertainty. Zhao et al[46] integrated the spectral information divergence and cosine for oil spill identification (light oil, Medium oil, lubricating oil and other oil) by airborne laser fluorescence radar, but the heavy fuel oil and crude oil need further recognizing. In recent years, local calibration modeling which combines the spectral search strategy with multivariate calibration method has been widely used, especially with the expansion of large-scale soil, feed and oil near infrared spectral library. Samples from different sources, dates and types aggravate the nonlinear relationship between spectra and concentration. To solve this problem, a group of samples, which were the most similar samples with the sample to be tested, were selected from the spectra database by a certain local modeling strategy to form to a calibration set[47]. In view of how to select the local samples and how to obtain the final prediction results, a variety of local modeling strategies were developed, such as CARNAC (Comparison Analysis using Restructured Near infrared And Constituent data) method, LWR (Locally Weighted Regression) method, and LOCAL method[48]. The red wine near infrared spectral library including 3000 samples established by Dambergs et al, the feed near infrared spectral library including 20000 samples established by Fernandez-Ahumada et al, and the soil near infrared spectral library including 1000 samples established by Genot et al were all used to predict key physical and chemical properties by local modeling strategy, respectively[49‒51]. The results were more accurate than the traditional modeling methods. This modeling strategy is not only used for nonlinear system, but also makes full use of advantage of spectrum database. It avoids updating models frequently by the traditional multivariate calibration methods, and is especially suitable for qualitative and quantitative analysis patterns of large-scale network spectral database. The method of spectral dimensionality reduction, which makes the search in the low and more characteristic dimensional space, is also a commonly used spectral searching strategy. Principal component analysis (PCA), isometric mapping (Isomap), locally linear embedding (LLE) and other statistical methods of dimensionality reduction are usually applied in supervised and unsupervised pattern recognition. However, fractal theory is paid more attention to spectral search[52]. For example, Lei et al[53] combined the wavelet transform with fractal theory to identify the near infrared spectra of lubricating oil. The fractal dimensions of the wavelet approximation and the detailed spectra were calculated, and the spectral searching was performed using those fractal dimensions as identification features, which achieved good results. 4 Conclusions In recent years, molecular spectroscopy has been getting more and more attention. For example, based on portable or handheld Raman and NIR instruments, the Division of Pharmaceutical Analysis (DPA) of FDA in USA is developing drug excipients spectral library to the samples polluted, adulterated and tampered in the drug production and supply chain. In China, it is also established the corresponding spectral database by the relevant departments in different fields gradually. The spectral searching method is one of the key technologies to make full use of these spectra database. This paper reviewed the existing searching algorithms and strategies. However, in practice, those should be improved according to specific objects, even renewed on the existing basis. Besides spectral searching algorithm and strategy, the experimental technique to build spectral library should be paid more attention to, so as to obtain high-quality (abundant information, strong characteristics, high SNR, and good repeatability) library spectra and unknown sample spectrum. The technique relies on equipment selection, sample pretreatment, accessories selection, parameter optimization, standardization of library process and other technical details. High-quality spectrum is the basis of all searching methods; therefore, to a large extent, the experimental technique to build spectral library is more important than searching method itself. Calibration Transfer is used to eliminate the variations between instruments in multivariate quantitative calibration[54], however the problem still exists in spectral searching technology. Differences exist not only between instruments of different brands, but also between instruments of the same model. With the wide application of the spectral library, this problem in spectral searching has obtained more and more CHU Xiao-Li et al. / Chinese Journal of Analytical Chemistry, 2014, 42(9): 1379–1386 attention[55‒59]. However, compared with multivariate quantitative calibration, the research about this question is not systematic, and few application reports are published, thus further research is needed. (1): 71‒73 [25] Wang Y, Ji L, Wang Y J, Liu X T, Niu Z Y, Du H Q. Engineering Plastics Application, 2005, 33(9): 47‒51 [26] Howari F M. J. Appl. Spectros., 2003, 70(5): 782‒787 [27] Reeves J B, Zapf C M. Appl. Spectros., 1999, 53(7): 836‒844 References [28] Chen T, Long X J, Wei L, Gong B, Li C M. Spectroscopy and [1] [29] He T, Shen J L. Spectroscopy and Spectral Analysis, 2013, Spectral Analysis, 2013, 33(2): 367‒370 Chu X L. Molecular Spectroscopy and Chemometrics, Beijing: Chemical Industry Press, 2011: 95‒108 [2] Blanco M, Romero M A. Analyst, 2001, 126(12) : 2212‒2217 [3] Terhoeven-Urselmans T, Vagen TG, Spaargaren O, Shepherd K D. Soil Sci. Soc. Am. J., 2010, 74(5): 1792‒1799 [4] Veij M D, Vandenabeele P, Beer T D, Remon J P, Moens L. J. Raman Spectros., 2008, 40(3): 297‒307 [5] Lavine B K, Fasasi A, Mirjankar N, White C, Mehta J. Microchemical Journal, 2014, 113(1) : 30‒35 33(9): 2348‒2353 [30] Guedes A, Ribeiro H, Fernández-González M, Aira M J, Abreu I. Talanta, 2014, 119(2): 473‒478 [31] van der Meer F. International Journal of Applied Earth Observation and Geoinformation, 2006, 8(1): 3–17 [32] Blanco M, Eustaquio A, González J M, Serrano D. Journal of Pharmaceutical and Biomedical Analysis, 2000, 22(1): 139‒148 [6] Zhou W H, Ying Y B, Xie L J. Applied Spectroscopy Reviews, [7] Chu X L, Yuan H F, Lu W Z. Progress in Chemistry, 2004, [8] Chang C I. IEEE Transactions on Information Theory, 2000, [9] Liu Y S, Cao M, Wang Y M, Luo G A. Chinese J. Anal. Chem., [36] Griffiths P R, Shao L M. Appl. Spectros., 2009, 63(8): 916–919 2006, 34(3): 333‒337 [37] Chu X L, Xu Y P, Tian S B, Wang J, Lu W Z. Chemometrics 2012, 47(8): 654‒670 16(4): 528‒542 46(5): 1927‒1932 [10] Zhang Q L, Wang X P, Zhao Y, Qu Meng, Du S Y. Chinese Journal of Spectroscopy Laboratory, 2013, 30(6): 2742‒2746 [11] Varmuza K, Karlovits M, Demuth W. Anal. Chim. Acta, 2003, 490(1-2): 313‒324 [12] Zhou W H, Xie L J, Ying Y B. Transactions of the Chinese Society of Agricultural Engineering, 2013, 29(19): 285‒292 [13] Meng Q H, Wang W B, Hu Y Z. China Journal of Chinese Materia Medica, 2007, 32(3) : 206‒210 [14] Tang T B, Yang H H, Liang X Z, Guo T, Li L Q, Luo G A. Journal of Guilin University of Electronic Technology, 2012, 32(5): 391‒397 [15] Li Y, Lü J W, Chen L. Pharmaceutical Care and Research, 2011, 11(4): 304‒306 [16] Khan S S, Madden M G. Chemometr. Intell. Lab., 2012, 114(1): 99‒108 [17] Plugge W, van der Vlies C. J. Pharm. Biomed. Anal., 1993, 11( 6) : 435‒442 [18] Ritchie G E, Mark H, Ciurczak E W. AAPS Pharm. Sci. Tech., 2003, 4(2): 109–118 [19] Feng Y C, Yang X L, Yang Z H. Journal of Chinese Pharmaceutical Sciences, 2011, 20(3): 290‒296 [20] Zhang X B, Yin L H. Chin. J. Pharm. Anal., 2011, 31(3): 603‒608 [21] Zhou W, Chen W. China Pharmacist, 2009, 12(4): 451‒452 [22] Nie L X, Wang G L, Li Z M, Lin R C. J. Infrared Millim. Waves, 2008, 27(3): 205‒209 [23] Tao Y, Dang L Z, Liu J, Liu Y, Li Q Q, Chen X. Chinese Journal of Spectroscopy Laboratory, 2013, 30(1): 27‒32 [24] Lu Y, Jiang L, Wu W W, Jin H. China Fiber Inspection, 2013, [33] Zhang X B, Ma J J, Cao L M. Chinese Journal of Spectroscopy Laboratory, 2013, 30(4): 2010‒2015 [34] Wang X L, Feng Y C, Hu C Q. Chinese J. Anal. Chem., 2009, 37(12): 1825‒1828 [35] Xu Y Q, Sun S Q, Xu J W. Chinese Journal of Spectroscopy Laboratory, 2002, 19(5): 606‒610 and Intelligent Laboratory Systems, 2011, 107(1): 44‒49 [38] Chu X L, Tian S B, Xu Y P, Wang J. Petroleum Processing and Petrochemicals, 2012, 43(1): 72‒77 [39] Li J Y, Chu X L, Tian S B, Lu W Z. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2013, 112(8): 457‒462 [40] Li J Y, Chu X L, Tian S B. Acta Petrolei Sinica (Petroleum Processing Section), 2013, 29(4): 655‒660 [41] Guo Z F, Dai L K. Chinese Journal of Spectroscopy Laboratory, 2013, 30(5): 2418‒2423 [42] Ramirez-Lopez L, Behrens T, .Schmidt K, Viscarra Rossel R A, Demattê J A M, Scholten T. Geoderma, 2013, 199(1): 43‒53 [43] Himmelsbach D S, Hellgeth J W, McAlister D D. J. Agric. Food Chem., 2006, 54(20): 7405‒7412 [44] Loudermilk J B, Himmelsbach D S, Barton F E, de Haseth J A. Appl. Spectros., 2008, 62(6): 661‒670 [45] Kong X B, Shu N, Tao J B. Spectroscopy and Spectral Analysis, 2011, 31(8): 2166‒2170 [46] Zhao C F, Qi M J, Ma Y J. China. Patent, CN 101923649 B, 2010 [47] Wei C L, Zhao Y G, Li D C, Zhang G L, Wu D W, Chen J K. Transactions of the Chinese Society of Agricultural Engineering, 2014, 30(1): 81‒88 [48] Shi X, Cai W S, Shao X G. Chinese Journal Anal. Chem., 2008, 36(8): 1093‒1096 [49] Genot V, Colinet G, Bock L, Vanvyve D, Reusen Y, Dardenne P. Journal of Near Infrared Spectroscopy, 2011, 19(2): 117‒138 [50] Dambergs R, Cozzolino D, Cynkar W, Gishen M. Journal of Near Infrared Spectroscopy, 2006, 14(1): 71‒79 CHU Xiao-Li et al. / Chinese Journal of Analytical Chemistry, 2014, 42(9): 1379–1386 [51] Fernandez-Ahumada E, Fearn T, Gomez-Cabrera A, Guerrero-Ginel J E, Perez-Marin D C, Garrido-Varo A. Appl. Spectros., 2013, 67(8): 924‒929 [52] Zhang P, Wang X K, C. Analyst, 1999,124(8): 1197‒1203 [56] Yoon W L, Jee R D, Moffat A C. Analyst, 2000, 125(10): 1817‒1822 Li H T, Li Y Y, Geng Y Z, Li F L. [57] Ma L Z, Guan L, Feng X L, Hu Z X, Xu J G, Bai Y S. Acta Chinese Journal of Quantum Electronics, 2007, 24(6): 672‒677 Petrolei Sinica (Petroleum Processing Section), 2013, 29(5): [53] Lei M, Feng X L. Journal of InstrumentalAnalysis, 2009, 28(5): 529‒534 [54] Chu X L, Yuan H F, Lu W Z. Spectroscopy and Spectral Analysis, 2001, 21(6): 881‒885 [55] Yoon W L, Jee R D, Moffat A C, Blackler P D, Yeung K, Lee D 891‒898 [58] Genot V, Colinet G, Dardene P, Bock L. Geophys. Res. Abstr., 2009, (11): 2805 [59] Lavine B K, Fasasi A, Mirjankar N, Sandercock M. Talanta, 2014, 119(2): 331‒340