Appl Biochem Biotechnol (2012) 166:711–721 DOI 10.1007/s12010-011-9460-3

Chemometric Analysis with Near-Infrared Spectroscopy for Chemically Pretreated Erianthus toward Efficient Bioethanol Production Yoshiki Horikawa & Tomoya Imai & Rie Takada & Takashi Watanabe & Keiji Takabe & Yoshinori Kobayashi & Junji Sugiyama

Received: 23 April 2011 / Accepted: 10 November 2011 / Published online: 30 November 2011 # Springer Science+Business Media, LLC 2011

Abstract In this paper, we report the combination of a near-infrared (NIR) spectroscopic method with multivariate analysis in order to develop a calibration model of the saccharification ratio of chemically pretreated Erianthus. The regression models clearly depend on the NIR spectral regions, and the information of CH and aromatic framework vibrations contributed most effectively to the alkaline dataset. From interpretations of the regression coefficient, lignin and cellulose were negatively and positively correlated with the saccharification ratio, respectively, and this result was supported by the data from wet chemical analysis. A more complex dataset was obtained from varied chemical pretreatments; here, the saccharification ratio was either small or had no linear correlation with each structural monocomponent. These results enabled the successful construction of the PLS regression model. NIR spectroscopy can be a rapid screening method for the saccharification ratio, and furthermore, can provide information of the key factors influencing the realization of more efficient enzymatic accessibility. Keywords Near-infrared (NIR) spectroscopy . Partial least-squares (PLS) regression . Regression coefficient . Saccharification ratio

Introduction The quantitative spectroscopy technique has been significantly modified and improved by using a variety of multivariate statistical methods. Regression is one such multivariate Y. Horikawa (*) : T. Imai : R. Takada : T. Watanabe : J. Sugiyama Research Institute for Sustainable Humanosphere, Kyoto University, Uji, Kyoto 611-0011, Japan e-mail: [email protected] K. Takabe Division of Forest and Biomaterials Science, Graduate School of Agriculture, Kyoto University, Kyoto, Kyoto, Japan Y. Kobayashi Tsukuba Research Laboratory, Japan Bioindustry Association, Tsukuba, Ibaragi, Japan

712

Appl Biochem Biotechnol (2012) 166:711–721

technique. In regression, two sets of variables are related—one or several dependent Yvariables are determined on the basis of independent X-variables, wherein X typically comprises large variables such as spectral data. Multiple regression (MLR) is the classical method to create a calibration model that directly correlates Y-variables with X-variables [1]. However, spectroscopic data are not applicable to MLR dealing with ordinal data because the X-variables themselves are typically highly correlated. Such a condition in which the independent variables themselves are correlated is called “multicollinearity,” and it leads to an almost singular matrix element and statistically unstable solutions. In order to overcome this problem, an indirect regression technique called the principal component regression (PCR) was developed. In PCR, the dimensionality of the data is first reduced, and subsequently, the principal components analysis (PCA) is applied to X [2, 3]. The PCA leads to a set of uncorrelated new variables (object scores) that can then be used in MLR in place of the original variables. Therefore, the PCR can be regarded as a two-step procedure: the PCA is initially used to transform X and the resulting novel matrix is then employed directly in the MLR model. An improvement to this method is the partial least-squares (PLS) regression model, alternatively this term was sometimes interpreted to projection to latent structure, that was reported by H. Wold [4]. This technique is partially common to the PCR because the principal components (PCs) are initially derived from X-variables, and the regression relation is then computed by combining the object scores with Y-variables. However, whereas the principal components in PCR are obtained only from the X-variables, in PLS, Y-variables are also used for the estimation of the object scores for X-variables; this provides a more high-quality information dataset of decreased dimensionality and eliminates data noise, leading to more accurate, robust, and reproducible calibration models. This sophisticated technique has been widely applied to quantitative analyses for more than two decades, for example, to the analysis of chromatographic [5], electrochemical [6], ultraviolet (UV) [7], infrared (IR) [8], and near-infrared (NIR) spectroscopic data [9, 10]. Recently, the excess usage of fossil fuels has increased research interest in renewable sources, and chemometric approaches combined with NIR spectroscopy have been reported for several biomass materials such as corn stover [11, 12], rice straw [13], and wheat straw [14]. Wolfrum and Sluiter [15] have reported the chemical compositions of corn stover after acid hydrolysis was employed as pretreatment and estimations were successfully conducted by employing NIR spectroscopy. We have also published the development of an acceptable calibration model of chemical pretreatments for rice straw on its structural components such as sugars and lignin as well as the saccharification ratio that is an important index to evaluate the pretreatments [16]. In this study, we report the establishment of the PLS calibration model of the saccharification ratio of alkaline pretreated Erianthus, which is well known as a desirable biomass material because of its high annual yields [17]. Further, we will discuss the factorial analysis to increase the saccharification potential by employing an interpretation of the regression coefficient. Furthermore, we will also describe corresponding attempts for more complicated datasets that were obtained by incorporating hydrothermal and dilute sulfuric acid pretreatment results into an alkaline dataset.

Materials and Methods Sample Preparation Erianthus was roughly milled by Orient Mill VM-16 (Seishin Enterprise Co., Ltd., Tokyo, Japan) and then finely milled by Bantam Mill AP-BL (Hosokawa Micron Corp., Osaka,

Appl Biochem Biotechnol (2012) 166:711–721

713

Japan) and passed through a 150-μm screen. The products were then subjected to various treatments: (1) alkaline treatment with sodium hydrate (0.1–1%) and ammonia water (5%) without cellulose crystalline transformation, (2) dilute sulfuric acid treatment (1%), and (3) hydrothermal treatment. These treatments were performed under different conditions of temperatures, reaction times, and chemical concentrations as parameters in order to vary the saccharification ratio widely. The data of sample preparations and the corresponding chemical analysis data will be published elsewhere (Kobayashi and Morikawa et al.). The residues after the chemical pretreatment were repeatedly washed in distilled water for chemical component analysis and enzymatic hydrolysis. Wet Chemical Analysis The sugars and lignin contents were measured by employing the standard protocol established by National Renewable Energy Laboratory (NREL) Laboratory Analytical Procedures [18]. The pretreated Erianthus was subjected to two-stage hydrolysis with sulfuric acid. The first treatment was conducted in 72% sulfuric acid at 30°C for 1 h, followed by a second treatment in 4% sulfuric acid at 121°C for 1 h.; this separated the monosaccharides in the liquor fraction from the solid components. This liquor also contained acid-soluble lignin that was not measured in this study. The monosaccharides that dissolved in hydrolysate liquor were measured by employing high-performance liquid chromatography (HPLC) using an HPLC system equipped with an Asahipak NH2P-50 4E column, an RF-AXL fluorescence detector, an autosampler, and a pneumatic controller (Prominence UFLC; Shimadzu Corp., Kyoto, Japan). The cellulose and hemicellulose contents (%) were estimated from the following equation: Cellulose ¼ 162=180  Glucose Hemicellulose ¼ ð162=180Þ  ðGalactose þ MannoseÞ þ ð132=150Þ  ðXylose þ ArabinoseÞ

The residue after the two-stage hydrolysis was burned in a furnace at 575°C for 5 h, and it was gravimetrically estimated to be ash content. The acid-insoluble lignin content was obtained by subtracting the ash content from the residue content. Enzymatic Hydrolysis The enzyme for the saccharification was a commercial cocktail named Accellerase 1500 (Genencor, Danisco USA, Inc., Rochester, NY). The washed solid fractions after pretreatment (100 mg of dry solids) were hydrolyzed at a 5% (wt./wt.) in 100 mM acetate buffer (pH 5.0) with the enzyme (40 filter paper activity (FPU)/g). The FPU was estimated in accordance with the standard procedure recommended by NREL [19]. The enzymatic hydrolysis was performed at 50°C for 24 h at a shaking speed of 150 rpm, and the sugar released was then determined as a reducing sugar by using the dinitrosalicylic acid method [20]. The saccharification ratio (%) is given by the following equation: Saccharification ratio ð%Þ ¼ ½fReducing sugar ðmg=mlÞ  2 ðmlÞ  0:9g=100 ðmgÞ  100

Near-Infrared Data Acquisition For reproducible measurements, a disk sample was molded by using a handpress after collecting approximately 0.04 g of the chemical pretreatments in accordance with a

714

Appl Biochem Biotechnol (2012) 166:711–721

published paper [16]. The NIR spectra were obtained on a PerkinElmer Spectrum 100N system equipped with a sphere diffuse reflectance accessory (PerkinElmer) and triglycine sulfate detector at a resolution of 16 cm−1 with the acquisition of 32 scans. The casted sample was directly placed on this accessory and recorded twice for monitoring the spectral repeatability. The original spectra were subjected to the second derivative by using the Savitzky–Golay algorithm [21] before the PCA and PLS regression analysis. Chemometric Analysis The PCA and PLS regressions applied to develop the calibration model of the saccharification ratio were generated by using a commercial software (Unscrambler® v.9.8; CAMO Software, Inc., Woodbridge, NJ). The optimum number of latent variables (LVs) used for PLS regression model was obtained from a full crossvalidation wherein a single sample was taken from the model and then predicted by using a model built without the sample. This process was repeated for every individual sample in the calibration set. The calibration models obtained were assessed by using the coefficient of determination (R2), the root-mean-square error of prediction (RMSEP) under full cross-validation (leave-one-out), and the ratio of performance to deviation (RPD), estimated by the ratio of the standard deviation of the reference data to the SEP.

Results and Discussion Construction of Partial Least-Squares Calibration Model for Alkaline Pretreatments The selection of the NIR region is essential for constructing a calibration model with high performances before the multivariate analysis. The full-length NIR range of 10,000– 4,000 cm−1 along with the various electromagnetic wave absorptions of the functional group from the biomass material were divided into four ranges in accordance with the properties of molecular vibrations (Fig. 1): (a) 10,000–7,300 cm−1, i.e., the range of the second or third overtone with little information from biomass material, (b) 7,300– 6,050 cm−1, corresponding to mainly OH overtone vibrations, (c) 6,050–5,500 cm−1, in which the CH vibrations and aromatic framework vibrations are present, and (d) 5,500– 4,000 cm−1, involving several combinational vibrations. The PLS regression analysis was performed by using each region and their combinations for the saccharification ratios estimated from several alkaline pretreatments by the wet chemical analysis, as shown in Table 1. In general, the optimum number of LVs used for the PLS regression is determined to minimizes the sum of squared residuals by using the cross-validation. Although the sum of the residuals decreases as the number of LVs increases without any inflection point, excess numbers may result in overfitting and make it an excessively complicated calibration model to understand the loading spectra or regression coefficients. Therefore, in the alkaline dataset, the number of LVs was made fewer than 3. With the exception of the model obtained for the NIR spectra ranging from 10,000 to 7,300 cm−1, all the PLS regression calibrations were successfully developed. The calibration performance for the range 6,050– 5,500 cm−1, involving RMSEP of 4.31 and R2 of 0.94, was better than that in the range 7,300–4,000 cm−1 wherein the corresponding values were 5.00 and 0.91, respectively. This result demonstrated that a broader range is not necessary to improve the model

Appl Biochem Biotechnol (2012) 166:711–721

715

Fig. 1 Second-derivative full NIR spectra separated from 4 regions: 10,000–7,300, 7,300–6,050, 6,050–5,500, and 5,500–4,000 cm−1. The black and gray lines are the alkaline pretreatments involving 14.08% and 90.89% of the saccharification ratio, respectively

1000

8000

6000

4000

Wavenumber(cm-1)

performance, and accordingly, the correct selection of the absorption range is a more important process to create a more reliable regression model. Interpretation of Calibration Model for Alkaline Pretreatments In order to find the key factor for increasing the saccharification ratio of alkaline treatments, the regression coefficient was examined in detail in the range of 7,300–5,500 cm−1 (Fig. 2); in this range, an acceptable calibration model as good as that for 6,050–5,500 cm−1 could

4000 cm-1

6050 cm-1 5500 cm-1

7300 cm-1

10000 cm-1

Table 1 Summary of enzymatic saccharification data and statistics of the calibration model for the alkaline pretreatments. A schematic image has been included on the left for assistance in understanding each spectral region

Saccharification ratio

Samples 45

Max. 90.89

Min. 14.08

Mean 63.65

10000 - 4000 cm -1 7300 - 4000 cm-1 7300 - 5500 cm-1 6050 - 4000 cm-1 10000 - 7300 cm -1 7300 - 6050 cm-1 6050 - 5500 cm-1 5500 - 4000 cm-1

LVs 2 2 2 2 2 2 2 2

RMSEP 5.01 5.00 4.32 4.95 15.66 6.22 4.31 5.13

R2 0.91 0.91 0.94 0.91 0.15 0.87 0.94 0.91

RPD 3.37 3.38 3.91 3.41 1.08 2.71 3.92 3.29

S.D. 16.97

SD standard deviation, LVs number of latent variables, RMSEP root-mean-square error of prediction, RPD ratio of performance to deviation

5794

5980

6287

6990

Fig. 2 a Enlarged spectra of Fig. 1 in the range of 7,300– 5,500 cm−1 and b regression coefficient for the calibration model obtained by using the corresponding region. The bold line at 5,980 cm−1 is assigned to lignin; the black dashed line at 5,794 cm−1 to hemicelluloses; the solid lines at 6,287 cm−1 and the series of bands around 6,770 and 6,670 cm−1 covered by the gray square were ascribed to cellulose; and finally, the gray dashed line at 6990 cm−1 is characteristic of amorphous regions in cellulose and hemicellulose

Appl Biochem Biotechnol (2012) 166:711–721 6770 6670

716

(a)

(b)

7300

6850

5950

6400

5500

-1

Wavenumber (cm )

be developed. In a regression model equation, the regression coefficients are the numerical coefficients that express the link between the saccharification ratio and NIR spectrum pattern. The bands of regression coefficient in the 6,050–5,500 cm−1 range were more completely intensified than those in the 7,300–6,050 cm−1 range, which indicates that the CH vibrations contributed to the construction of the calibration model compared to the OH absorption. The assignment of the bands for the lignocellulosic biomass in these regions has been reported by Tsuchikawa, Siesler, and co-workers [22–24]; this is helpful for interpreting the PLS calibration model. The most remarkable band in this coefficient regression is a positive absorption at 5,980 cm−1 that has been assigned to the aromatic skeleton from lignin. Since the PLS regression was obtained from the second derivative spectra, the positive band at 5,980 cm−1 indicates that the lignin contents should have a negative correlation with the saccharification ratio. The 5,794 cm−1 band has been assigned to furanose or pyranose that is present due to hemicellulose. For this band, the regression coefficient is observed to be positive; this implies that the hemicelluloses in alkaline pretreated Erianthus may have a negative relation with the saccharification ratio. For the OH overtone region, the band at 6,990 cm−1 appears to be a mixture of the OH information from hemicellulose and cellulose from disordered area; in this case, these polysaccharides might have a negative correlation with the saccharification ratio. At lower frequencies, a series of bands at around 6,700 cm−1 and the band at 6,287 cm−1 were assigned to semicrystalline and crystalline regions in cellulose microfibrils. In contrast to lignin and hemicellulose, these bands related to cellulose showed negative regression coefficients for absorptions; this suggests that an increase in the cellulose content should lead to higher values of the saccharification ratio. To confirm this suggestion that is based on the regression coefficient, the major structural components in Erianthus were estimated by wet

Appl Biochem Biotechnol (2012) 166:711–721

717

chemical analysis. The relationships between the saccharification ratio and three components are presented in Fig. 3; similar to the above argument, cellulose and lignin showed positive and negative linear correlations, respectively. Moreover, although there is variability, hemicellulose was observed to tend to be negatively correlated with the saccharification ratio, which agrees well with the interpretation of the regression coefficient. Previous studies have reported that alkaline pretreatment conducted on the cellulose microfibril swelling leads to reductions in the degree of polymerization, structural cleavages between lignin and carbohydrates, and disorganization of the lignin structure; as a result, the internal surface area for susceptibility increases [25, 26]. In the case of chemical contents, alkaline pretreatment preferentially removed lignin from lignocellulosic biomass with increase of saccharification ratio, harsh treatment further effected to susceptible polysaccharide such as hemicellulose, and subsequently crystalline cellulose remained, which is consistent with the inspection of regression coefficient. For the negative correlation between saccharification ratio and hemicellulose contents, the reason may be that the enzymatic cocktail used in this study is less activity for xylan or other hemicelluloses. Then, hemicellulose remained on the cellulose even during enzymatic hydrolysis and inhibited the accessibility of cellulase. PCA Analysis for Various Pretreated Samples

Saccharification ratio (%)

We now attempt to develop a calibration model for more complex datasets. We have conducted dilute sulfuric acid and hydrothermal treatments whose processes also enhance the enzymatic accessibility; however, the behaviors are not similar to the alkaline treatment in that the hemicellulose is preferentially attacked [27], and moreover, the residues among this mixture sample set are fairly different. In order to ensure variability in the alkaline and additional samples, PCA was performed in the wavelength region of 7,300–5,500 cm−1 (Fig. 4) wherein higher calibration performances have been demonstrated in the alkaline dataset. In the score plots, the alkaline samples were localized in the first and fourth quadrants whereas the samples from dilute sulfuric acid treatments or hydrothermal treatments were mainly distributed in the second or third quadrant. The PC1 loading spectra from PCA exhibited positive bands at 5,980 cm−1, which implies a decrease in the lignin contents and shifts to higher scores for PC1; this corresponds to the position of the alkaline samples in the first and fourth quadrants. The converse was also true because the dilute sulfuric acid and hydrothermal treatments preferentially removed the hemicellulose and

(a)

(b)

(c)

100

100

100

R2 = 80 0.88

80

80

60

60

60

40

40

40

20

20

20

0 30

40

50

60

Cellulose content (%)

70

0 15

20

25

30

Hemicellulose content (%)

0

R2 = 0.72

0

5

10

15

20

Lignin content (%)

Fig. 3 Relationships of with saccharification ratio and the three major structural components a cellulose, b hemicelluloses, and c lignin, for the alkaline dataset measured by wet chemical analysis

718

Appl Biochem Biotechnol (2012) 166:711–721

Fig. 4 a Principal components analysis (PCA) score plotted on the first and second principal components of near-infrared (NIR) spectra recorded from 58 samples containing alkaline (open circle), hydrothermal (filled triangle), and dilute sulfuric acid pretreatments (filled square). b Black and gray lines are PC1 and PC2 loading in PCA. The bands at 6,990 and 5,980 cm−1 are assigned to the amorphous regions in cellulose and hemicellulose or lignin, respectively

(a) 10

PC2 (14%)

5

0

-5

-10 -20

-10

0

10

7300

6990

(b)

5980

PC1 (72%)

6850

6400

5950

5500

Wavenumber(cm-1)

consequently retained the lignin. In the PC2 loading spectra, the positive band at 6,990 cm−1 has been assigned to susceptible polysaccharides. Therefore, for the samples subjected to dilute sulfuric acid, the hemicelluloses and accessible cellulose may be removed to a greater extent than that in the case of the hydrothermal treatment. Construction of PLS Calibration Model for Various Pretreatments In addition to Fig. 3, the relationships between the saccharification ratio and structural components are presented in Fig. 5. In Fig. 5a, even though the cellulose content seems to have a slightly positive relation, the correlation of each structural component with the saccharification ratios in various datasets was lower than that in alkaline data. This is indicative of the difficulties in predicting the saccharification ratio by employing single structural components. The PLS regression analysis has been performed on a variety of treated datasets which need to larger number of LVs, so we developed regression model for each region with LVs less than 5; the summary statistics are shown in Table 2. Surprisingly, these models, except

Saccharification ratio (%)

Appl Biochem Biotechnol (2012) 166:711–721

719

(a)

(b)

(c)

100

100

100

80

80

80

60

60

60

40

40

40

20

20

20

0 30

0 40

50

60

70

Cellulose content (%)

0 0

10

20

30

Hemicellulose content (%)

0

10

20

30

40

Lignin content (%)

Fig. 5 Relationships of the saccharification ratio with the major structural components a cellulose, b hemicelluloses, and c lignin, for a variety of datasets of alkaline (open circle), hydrothermal (filled triangle), and dilute sulfuric acid pretreatments (filled square), measured by wet chemical analysis

for the one constructed by the NIR region in the 10,000–7,300 cm−1 range, exhibited higher performances when R2 >0.89, RMSEP<6.34, and RPD>2.98. The best accuracy was obtained from the spectrum ranging from 7,300 to 5,500 cm−1, which indicates that the OH and CH vibration strongly contributed to generate the calibration model. Therefore, the regression coefficient in this region was computed for realizing the model and performing comparisons with the alkaline dataset (Fig. 6). The regression coefficient patterns from various pretreatments were similar to the alkaline ones; the strong lignin band at 5980 cm−1 exhibited positive correlations while a series of cellulose bands were negatively correlated. In the case of hemicelluloses, the band intensity at 5,794 cm−1 was significantly negatively correlated with the enzymatic hydrolysis after the addition of dilute sulfuric acid and the hydrothermal sample. Interestingly, the 6,900 cm−1 band in which a shoulder absorption is located in the second-derivative spectra of the alkaline pretreatments was clearly visible in the dilute sulfuric acid and hydrothermal treatments. This band showed a negative regression coefficient, which suggests a positive relation with the enzymatic hydrolysis. For the dilute sulfuric acid and hydrothermal treatments, the saccharification ratio was positive correlated with cellulose and lignin (Fig. 6a and c). If the band at 6,900 cm−1 is

4000 cm -1

6050 cm -1 5500 cm -1

7300 cm -1

10000 cm -1

Table 2 Summary of enzymatic saccharification data and statistics of the calibration model for the varied pretreatments. A schematic image has been included on the left for assistance in understanding each spectral region

Saccharification ratio

Samples 58

Max. 90.89

Min. 12.89

Mean 58.36

10000 - 4000 cm -1 7300 - 4000 cm -1 7300 - 5500 cm -1 6050 - 4000 cm -1 10000 - 7300 cm -1 7300 - 6050 cm -1 6050 - 5500 cm -1 5500 - 4000 cm -1

LVs 4 4 4 4 4 4 4 4

RMSEP 6.06 6.09 5.09 6.14 14.81 5.81 5.77 6.33

R2 0.90 0.90 0.93 0.89 0.38 0.90 0.91 0.89

RPD 3.10 3.08 3.69 3.06 1.27 3.23 3.26 2.97

S.D. 18.85

SD standard deviation, LVs number of latent variables, RMSEP root-mean-square error of prediction, RPD ratio of performance to deviation

5794

5980

6287

Fig. 6 a The second-derivative spectra (7,300–5,500 cm−1) of alkaline, hydrothermal and dilute sulfuric acid pretreatments; their saccharification ratios were 48.9%, 50.6%, and 50.3%. b Regression coefficient for the calibration model obtained from the variously pretreated dataset. Bold lines at 6,990 and 5,980 cm−1 would be assigned to lignin; the black dashed line at 5,794 cm−1 to hemicelluloses; and the solid lines at 6,287 cm−1 and the series of bands around 6,770 and 6,670 cm−1 covered by the gray square to cellulose. The gray dashed line at 6,990 cm−1 is characteristic of amorphous regions in cellulose and hemicellulose

Appl Biochem Biotechnol (2012) 166:711–721

6990 6900 6770 6670

720

(a)

(b)

7300

6850 6400 5950 Wavenumber (cm-1)

5500

characteristic of cellulose, it can be expected to appear in the regression coefficient of the alkaline dataset, and hence, we would ascribe the 6,900 cm−1 band to the OH vibration due to lignin, the assignment of which is corresponding to the published paper [28].

Conclusion In this paper, we have successfully demonstrated the application of NIR spectroscopy with the chemometric technique for establishing the calibration model of the saccharification ratio of alkaline pretreated Erianthus. The development of the PLS regression model from a variety of NIR region clarified the significant contributions of the CH and aromatic framework vibrations. The regression coefficient can be used to illustrate the important factors for constructing the calibration model since it shows negative and positive correlations of the saccharification ratio with lignin and cellulose, respectively. In this paper, this interpretation was elucidated by quantitative measurements of the major structural components through wet chemical analysis. Dilute sulfuric acid and hydrothermal treatments were conducted on alkaline pretreated Erianthus to prepare more complicated datasets in which the saccharification ratio has slight or no linear correlation with each structural monocomponent. Nevertheless, the NIR spectroscopy enabled the creation of an acceptable calibration model. In this manner, we have demonstrated that the NIR/ chemometric analysis has great potential as a rapid and precise screening method as well as for assisting the elucidation of the mechanism of enzymatic saccharification under chemical pretreatments.

Appl Biochem Biotechnol (2012) 166:711–721

721

Acknowledgments This study is supported by the New Energy and Industrial Technology Development Organization (NEDO). The authors would like to thank Makiko Imai and Keiko Kanai for their experimental discussion and assistance.

References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28.

Haaland, D. M., Easterling, R. G., & Vopicka, D. A. (1985). Applied Spectroscopy, 39, 73–83. Fredericks, P. M., Lee, J. B., Osborn, P. R., & Swinkels, D. A. J. (1985). Applied Spectroscopy, 39, 303–310. Fredericks, P. M., Lee, J. B., Osborn, P. R., & Swinkels, D. A. J. (1985). Applied Spectroscopy, 39, 311–316. Wold, H. (1974). European Economic Review, 5, 67–86. Lindberg, W., Ohman, J., Wold, S., & Martens, H. (1985). Analytica Chimica Acta, 171, 1–11. Otto, M., & Thomas, J. D. R. (1985). Analytical Chemistry, 57, 2647–2651. Lindberg, W., Persson, J. A., & Wold, S. (1983). Analytical Chemistry, 55, 643–648. Holmgren, A., & Norden, B. (1988). Applied Spectroscopy, 42, 255–262. Geladi, P., Macdougall, D., & Martens, H. (1985). Applied Spectroscopy, 39, 491–500. Martens, M., & Martens, H. (1986). Applied Spectroscopy, 40, 303–310. Sanderson, M. A., Agblevor, F., Collins, M., & Johnson, D. K. (1996). Biomass and Bioenergy, 11, 365–370. Hames, B. R., Thomas, S. R., Sluiter, A. D., Roth, C. J., & Templeton, D. W. (2003). Applied Biochemistry and Biotechnology, 105, 5–16. Jin, S. Y., & Chen, H. Z. (2007). Industrial Crops and Products, 26, 207–211. Bruun, S., Jensen, J. W., Magid, J., Lindedam, J., & Engelsen, S. B. (2010). Industrial Crops and Products, 31, 321–326. Wolfrum, E. J., & Sluiter, A. D. (2009). Cellulose, 16, 567–576. Horikawa, Y., Imai, T., Takada, R., Watanabe, T., Takabe, K., Kobayashi, Y., et al. (2011). Applied Biochemistry and Biotechnology, 164, 194–203. Mislevy, P., Kalmbacher, R. S., Overman, A. J., & Martin, F. G. (1986). Biomass, 11, 243–253. Sluiter, A., Hames, B., Ruiz, R., Scarlata, C., Sluiter, J., Templeton, D. & Crocker, D. (2008). National renewable energy laboratory. Golden, CO, USA. from http://www.nrel.gov/biomass/analytical_procedures.html. Adney, B. & Baker, J. (1996). National renewable energy laboratory. Golden, CO, USA. from http:// www.nrel.gov/biomass/analytical_procedures.html. Luchsinger, W. W., & Cornesky, R. A. (1962). Analytical Biochemistry, 4, 346–347. Savitzky, A., & Golay, M. J. E. (1964). Analytical Chemistry, 36, 1627–1639. Tsuchikawa, S., & Siesler, H. W. (2003). Applied Spectroscopy, 57, 667–674. Tsuchikawa, S., & Siesler, H. W. (2003). Applied Spectroscopy, 57, 675–681. Tsuchikawa, S., Yonenobu, H., & Siesler, H. W. (2005). Analyst, 130, 379–384. Fan, L. T., Gharpuray, M. M., & Lee, Y.-H. (1987). Cellulose hydrolysis. Berlin: Springer. Sun, Y., & Cheng, J. Y. (2002). Bioresource Technology, 83, 1–11. Hendriks, A. T. W. M., & Zeeman, G. (2009). Bioresource Technology, 100, 10–18. Mitsui, K., Inagaki, T., & Tsuchikawa, S. (2008). Biomacromolecules, 9, 286–288.