PREDICTION of WOOD SPECIES and PULP BRIGHTNESS from ROUNDWOOD MEASUREMENTS by David Nilsson

PREDICTION OF WOOD SPECIES AND

PULP BRIGHTNESS FROM

ROUNDWOOD MEASUREMENTS

David Nilsson

Akademisk avhandling

Som med tillstånd av rektorsämbetet vid Umeå universitet för erhållande av Filosofie Doktorsexamen vid Teknisk–Naturvetenskapliga fakulteten, framlägges till offentlig granskning vid Kemiska institutionen, Umeå universitet, sal KB3A9, KBC, fredagen den 28:e oktober, kl 10.00, 2005.

Fakultetsopponent: Prof. Olav Kvalheim, Department of Chemistry, University of Bergen, Allegaten 41, N-5007 Bergen, Norway.

-1-

-2- Title Prediction of Wood Species and Pulp Brightness from Roundwood Measurements

Author David Nilsson, Department of Chemistry, Organic Chemistry, Umeå University, SE–901 87, Umeå, Sweden.

Abstract This thesis presents a number of studies, where a multivariate approach was taken to construct models that predict wood species and thermo mechanical pulp brightness from roundwood of Norway spruce and Scots pine. The first and second studies produced multivariate prediction models for wood species from the bark of spruce and pine. These models can be used for wood species classification and would replace the manual log assessment that takes place today. Principal Component Analysis, PCA, and Partial least squares projections to Latent Structures, PLS, were used to predict the wood species from multivariate measurements recorded from the bark of spruce and pine. Two different kinds of measurements were employed, near-infrared spectroscopy and digital imaging. Both methods showed that it was possible to predict the wood species with a high accuracy. The third and fourth studies of the thesis are related to the wood storage of roundwood and the deterioration of wood that occurs during the storage. The third study used an experimental design with five storage factors that provided different conditions for the analysed wood. The experimental design made it possible to identify the factors and the interaction between factors, which were important for the ISO brightness of peroxide and dithionite bleached thermo mechanical pulp, TMP. The final study of the thesis used NIR spectroscopy for predicting the ISO brightness of bleached TMP. Spectra recorded from stored wood were used to construct PLS prediction models.

Keywords Design of Experiments, ISO brightness, Multivariate Image Analysis, Multivariate Modelling, NIR spectroscopy, Norway spruce, PCA, PLS, Roundwood, Scots pine.

ISBN: 91-7305-959-5

-3-

-4- Contents

1. List of Papers………………………………………………7 2. List of Abbreviations…………………………….……….. 8 3. Introduction…………………………………..…………… 9 4. Design of Experiments……………………………………. 13 5. Multivariate Modelling………………………………...... 19 5.1. Data Pre-treatment…………………………………. 20 5.2. Principal Component Analysis, PCA………………. 21 5.3. Partial least squares projections to Latent Structures, PLS……………………………... 24 5.4. Model Complexity and Validation…………. ……... 26 6. NIR Spectroscopy………………………………………….29 7. Modelling of Multivariate Images………………….……. 35 8. Roundwood Wood Species Classification……………….. 41 8.1. Roundwood species classification using NIR spectroscopy, Paper I…………………………… 41 8.2. Roundwood species classification using Multivariate Image Analysis, Paper II………………. 43 8.3. Comparison between the two methods…………….. 46 9. Wood Storage and its Relation to the Brightness of Bleached Thermo Mechanical Pulp……… 49 9.1. The Effect of Designed Wood storage on the Brightness of Bleached TMP, Paper III…………...... 51 9.2. Prediction of Bleached TMP Brightness using NIR Spectroscopy on Wood Raw Material, Paper IV……... 56 10. Concluding Remarks……………………………………... 59 11. References…………………………………………………. 61 12. Acknowledgements………………………………………...71

-5-

-6- List of Papers

1. List of Papers

This thesis is based on the papers listed below, which will be referred to in the text by the corresponding Roman numerals (I-IV).

I Nilsson, D., E. Persson, Classification of the Softwood Species Spruce and Pine using Near-infrared Reflectance Measurements on Bark combined with Multivariate Data Analysis, Submitted to Wood Science and Technology

II Nilsson, D., U. Edlund, “Pine and Spruce Roundwood Species Classification Using Multivariate Image Analysis on Bark”, to be published in Holzforschung. 59(6) (2005). Reprinted with kind permission from Walter de Gruyter, GmbH & Co KG, Berlin Germany.

III Nilsson, D., U. Edlund, K. Elg-Christofferson, M. Sjöström, R. Agnemo (2003). “The effect of designed wood storage on the brightness of bleached and unbleached thermo mechanical pulp.” Nordic Pulp & Paper Research Journal. 18(4): 369-376. Reprinted with kind permission from Nordic Pulp and Paper Research Journal, Stockholm, Sweden.

IV Nilsson, D., U. Edlund, M. Sjöström, R. Agnemo (2005). “Prediction of thermo mechanical pulp brightness using NIR spectroscopy on wood raw material.” Paperi Ja Puu – Paper and Timber. 87(2): 102-109. Reprinted with kind permission from Paperi Ja Puu Oy, Helsinki, Finland.

-7- List of Abbreviations

2. List of Abbreviations

Abbreviation Meaning DFT Discrete Fourier Transformation DWT Discrete Wavelet Transformation GLCM Grey Level Co-occurrence Matrix MIA Multivariate Image Analysis MLR Multiple Linear Regression MSC Multiplicative Scatter Correction NIR Near InfraRed (Reflectance) O-PLS Orthogonal PLS (see PLS) OSC Orthogonal Signal Correction PCA Principal Component Analysis PLS Partial least squares projections to Latent Structures PLS-DA PLS Discriminant Analysis RGB Red Green Blue RMSEP Root Mean Squared Error of Prediction TMP Thermo Mechanical Pulp

-8- Introduction

3. Introduction

hermo mechanical pulp, TMP, is a high-yield pulp Tmanufactured in many countries, among which Canada, USA, Finland, Sweden, Japan and Norway are the largest producers. The difference between the thermo mechanical pulping process and the mechanical one is that the former heats the wood chips with steamed water prior to pulping. Unlike the cooked chemical sulphite and sulphate pulps, the mechanical pulping process refines wood into pulp that, with the exception of bark, consists of all wood raw material components including lignin. Lignin is a complex polymer that can be found in all types of wood where it serves as a binding material between the wood fibres. Lignin will react with the ultraviolet light of the sun and this will cause the paper produced from the TMP to yellow. Therefore, TMP is primarily used for low-cost and non-durable papers, i.e. in most cases newspaper.

In Sweden TMP is produced from the two softwood species Norway spruce (Picea Abies) and Scots pine (Pinus Sylvestris). These are harvested and then transported to saw and pulp mills. Pulpwood logs are normally of poorer quality and lower diameter than sawlogs. Regardless of the quality of the used roundwood, tree harvesting and transportation to the mills will imply that the logs are stored at several locations before they enter the actual manufacturing process. Normally, the felled trees are stored temporarily at the roadside prior to transport to the saw or pulp mills. At the mills, the logs are then stored for a longer time to ensure a shortage-free production. The storage sometimes brings unwanted effects to the quality of the wood, especially at summertime when for example insect attacks and a high temperature can deteriorate the stored logs. Stored roundwood is generally sprinkled with water during the summer in order to keep the wood from drying. Consequently, sprinkling has a positive effect on the quality of sawlogs, but it may also have undesirable effects on the brightness of the stored wood. In TMP production, a darker wood will have a negative effect on the bleachability of the pulp.

-9- Introduction

At the sawmill, the logs are manually classified as either spruce or pine prior to sawing. The residual parts of the logs not utilised as sawn wood are chopped into wood chips and transported to pulp mills. Thus, the wood used for TMP enters the process as either pulpwood or saw mill chips. The pulpwood logs are debarked and chopped into chips. After steaming the chips with water, the wood is refined into TMP. The pulp is then bleached in order to achieve acceptable brightness. This is normally done by treating the TMP with hydrogen peroxide or sodium dithionite. The bleaching can consist of a single or of two and even more stages. The bleachability of the pulp will be determined directly by the quality of the wood used. Wood of lesser quality may prevent the pulp from achieving acceptable brightness and may increase the amount of bleaching chemicals that has to be used.

With access to powerful personal computers, wood and pulp have, in the last decade, been subject to a large number of investigations that characterise the material from multivariate data, i.e. data comprised of many variables. Analysing multivariate of data with traditional univariate statistical methods is troublesome. The sheer number of variables and also variable collinearity, i.e. several variables that give the same information, cause problems that are difficult to handle. Principal Component Analysis, PCA (Wold et al. 1987), and Partial least squares projections to Latent Structures, PLS (Sjöström et al. 1983; Geladi and Kowalski 1986), are two multivariate data analysis methods that compress the data into a few significant variables. These methods also deal with the collinearity problem. By using these methods, it is possible to construct models that can be used to predict the properties of the analysed material. The information content in the models might be increased by recording the data from a set of designed samples. Design of experiments is a way to set up an experiment in such a way that the information is maximised.

This thesis presents results that are based on experimental data from two different analysis methods that generate multivariate data. The first is Near-Infrared Reflectance (NIR) spectroscopy (Osbourne et al. 1993), which is a fast analysis method that is used for quality control and monitoring of variables in various processes of today. The second method is Multivariate Image Analysis, MIA (Geladi et al 1989). This method treats a digital image in terms of multiple variables that can be processed by means of multivariate data analysis. The above mentioned methods are here used for recording data from logs of roundwood. The first property that is correlated from the data is the wood species, i.e. pine or spruce.

-10- Introduction

This property represents an early stage of the refinement of wood. The second property that is modelled from the data is the ISO brightness of bleached TMP. This property, on the other hand, is characteristic of the final paper product.

The first aim of the thesis is to demonstrate how it is possible to construct models from roundwood measurements that predict the wood species and the ISO brightness. The second aim of the thesis is to show how design of experiments can be used to identify factors that affect the bleachability of TMP. The structure of the thesis will be the following; the fourth chapter will explain the concept of design of experiments and why its use is important. In chapter five to seven, the reader will in be familiarised with multivariate data analysis followed by NIR spectroscopy and multivariate image analysis. The last chapters, eight and nine, will deal with wood species classification and wood storage. Finally, there will be some concluding remarks.

-11-

-12- Design of Experiments

4. Design of Experiments

esign of experiments involves a set of variables or factors that D are varied in such a way that the information that can be gathered is maximised. It should be understood that an experimental design comprises a number of experiments followed by the subsequent measuring of the outcome, where all factors are varied simultaneously in such a way that each of the experiments represents a set of factor settings that is not matched by any of the other experiments. This implies that the number of experiments needed is kept at a minimum. A designed experiment will always span the experimental samples more efficiently than if they are just picked randomly. With the use of an experimental design it is possible to determine interaction effects between the factors, making this approach far better than changing-one-variable at-a-time methodology, which is probably the most spontaneous way of performing experiments. To illustrate the difference between the two approaches, the model for dithionite-bleached TMP samples in Paper III can be used as an example. Figure 1 shows a contour plot for the decrease in ISO brightness for TMP produced from wood that has been stored for 14 weeks. Using the changing-one-variable-at-a-time approach, varying the light according to the horizontal line and then the watering according to the vertical line will eventually give the factor settings that correspond to the largest decrease in ISO brightness. Figure 2 shows the ISO brightness decrease when the response is also depending on the interaction between light and watering. It is clear that the approach of changing-one-variable- at-a-time will not give the settings that correspond to the largest decrease in ISO brightness. However, figure 3 shows how design of experiments can be used for varying the two factors at the same time. Obviously, this approach will have a better chance of giving the factor settings that provide the largest decrease in ISO brightness. The factors of light and watering are here varied in a full factorial design at two levels

-13- Design of Experiments

Figure 1. Contour plot from the dithionite model in Paper III. Changing the light and then the watering will eventually give the factor settings that correspond to

the largest decrease in ISO brightness.

Figure 2. But if there exists an interaction, changing one variable at a time will not lead the experiments into the right direction.

Figure 3. Instead, it is better to change both variables at a time as this approach will have a better chance of reaching settings that corresponds to the largest

decrease in ISO brightness.

-14- Design of Experiments

(Lundstedt et al. 1998). The number of experiments is given by nk, where n equals the number of levels and k is the number of variables or factors. In this example, both the number of levels and the number of factors is equal to two which equals 22 = 4 experiments. A design with 2 levels and 3 factors would be denoted 23 and equal 8 experiments. At two levels the factors are varied between the low setting usually denoted (-) and the high setting denoted (+). Full factorial design tables for two and three factors can be seen in table 1 and table 2. Also, a number of centre points are normally added to the designs in order to estimate the replicate error and give indications of model curvature. The centre level setting is usually denoted as (0).

Exp. No x1 x2 1 + + 2 - + 3 + - 4 - - Table 1. The table for the full factorial design in two levels when the two factors x1 and x2 are varied in a total of four experimental runs.

Exp. No x1 x2 x3 1 + + + 2 - + + 3 + - + 4 - - + 5 + + - 6 - + - 7 - - 8 + - - Table 2. The table for the full factorial design in two levels when the three factors, x1, x2 and x3 are varied in a total of eight experimental runs.

A function y = f(x) can then be fitted between the measured response and the factors. Multiple Linear Regression (MLR) or PLS can be used for fitting the model. The simplest model for two factors will be the linear model

y = b0 + b1x1 + b2 x2 + f

-15- Design of Experiments

where b0 is a constant, b1 and b2 the coefficients for the two factors x1 and x2 and f is the residual. Normally, there is a desire to determine interactions between the factors which gives the model

y = b0 + b1 x1 + b2 x2 + b12 x1 x2 + f where b12 is the coefficient that describes the interaction between the two factors.

The number of possible experiments that can be carried out is a limiting factor in an investigation. The normal approach is that after a couple of preliminary trial experiments are performed, a screening design is set up in the purpose of getting a rough estimation of the measured response and the effect of the factors. A design can be constructed for optimisation purposes or just for examination how different factors affect the response. As the number of experiments required rises exponentially with the number of factors, it can be necessary to reduce a full factorial design to a fractional factorial design. This is done by generating the settings for one or more factors from interaction effects of the other factors. The main drawback of doing this is that these interactions cannot be resolved as they are confounded with the main factors. The least important interactions should therefore be discarded in this way. An example to this is the 24-1 fractional factorial design that requires eight experiments compared to 16 experiments for the corresponding full factorial design. The fourth factor is generated here from the interaction between the first three factors. This interaction is consequently lost, but all main factor coefficients and also some of the two-factor interactions can be resolved. The reader is referred to the following literature (Box and Draper 1987; Myers and Montgomery 1995; Montgomery 1997; Lundstedt et al. 1998) for a deeper understanding how fractional factorial designs can be generated and how additional trials in the experimental region can be constructed after a screening design has been performed.

In addition to the full and fractional factorial designs there are other ways to construct an experimental design. The D-optimal design (Myers and Montgomery 1995) is a computer generated design that chooses the experimental runs from a set of possible experiments. The selection is made on basis of the size of the determinant of the chosen design matrix X. The use of a D-optimal design can be beneficial when the examined

-16- Design of Experiments factors are of a qualitative nature with several levels. The use of D- optimality will in this case give a more efficient design with respect to the number of experimental runs compared to a factorial design. In this thesis, the design used for the wood storage discussed in Paper III was D- optimal. As several of the factors were quantitative, this type of design made it possible to carry out the experiments in the limited number of climate chambers that were available for the storage.

-17-

-18- Multivariate Modelling

5. Multivariate Modelling

typical scientific study often generates large amounts of data A that is gathered in a matrix, X. This matrix is comprised of N observations (e.g. wood samples) as rows and K variables (e.g. NIR wavelengths) as columns (see figure 4). The size of the matrix is dependent on the scope of the study and how many variables are measured, but matrices exceeding a number of one thousand for the variables are common. The number of observations is, for most studies, more limited. Analysing a large matrix is problematic in several aspects and two questions need to be answered. First of all, how is it possible to gather information from so many variables, and secondly, what happens if some of the variables are correlated? It is obvious that neither of these two questions can be answered by means of traditional univariate statistical analysis where one variable is analysed at a time. Instead, it is better to use a multivariate method that analyses the data set with respect to all variables at the same time. Principal Component Analysis (Joliffe 1986; Wold et al. 1987), PCA, and Partial least squares projections to Latent Structures (Sjöström et al. 1983; Geladi and Kowalski 1986), PLS, are two methods used in multivariate analysis. The concept behind these methods is that they compress matrices into a few latent variables that explain most of the variation that can be found in the data. The calculation of PCA and PLS models makes it much simpler to obtain an overview of data and at the same time solves the problem with correlated variables.

K Figure 4. The typical scientific study often generates data that can be assembled in a matrix, X. The rows, N, represent the number of X studied samples and the columns, K, are the same as the measured

variables.

-19- Multivariate Modelling

5.1 Data Pre-treament

he procedure for how multivariate methods are applied to data T involves a data pre-treatment step where the matrix is processed in such a way that it facilitates the modelling. The most common data pre-treatment is variable mean centering, which implies that the variable means are deducted from the variable values. Visually this means that the matrix is shifted towards the origin of coordinates. It is praxis that variable mean centering is always performed prior to PCA or PLS modelling. Sometimes variable scaling is also applied to the data set. The most frequently used scaling is unit-variance (UV) scaling where each variable is divided by its standard deviation. Scaling into unit variances implies that all variables have the same kind of variation and they therefore are equally important. For certain types of data (e.g. non spectral) this is essential for successful modelling but for other kinds of data (e.g. spectral data) unit variance scaling will lessen the signal-to- noise ratio which will negatively affect the multivariate modelling.

In addition to different scaling techniques, a matrix can be pre-treated by means of applying a filter that removes unwanted information. Multiplicative Scatter Correction (Geladi et al. 1985), MSC, and Orthogonal Signal Correction (Wold et al. 1998), OSC, are two methods which are used for filtering X data. In the present thesis, these two methods have been applied to NIR data and this will be discussed more thoroughly in the next chapter.

-20- Multivariate Modelling

5.2 Principal Component Analysis, PCA

CA is a multivariate method that is applied to a data set or P matrix, X, in order to reduce the dimensionality of the data. The PCA algorithm calculates a set of principal components that are orthogonal to each other and explains most of the variation in the data. Each of the components is comprised of a score vector, t, and a loading vector, p. The first principal component can be seen as a line that goes through the origin of coordinates and is drawn through the K-dimensional space in such a way that it covers most of the variation of the data. Each observation is then projected down to the line resulting in a score value can be obtained for all observations. The next principal component also goes through the origin and is orthogonal to first component. In the same way, each observation is projected down to the line so that a score vector for the second component can be obtained. Here, the first and second components form a plane in the K-dimensional space. The loading vectors for the two components are derived from the way in which this plane is located in the K-dimensional space in relation to the original variables. Thus, the length of each score vector is equal to the number of observations while the length of each loading vector equals the number of variables.

The PCA decomposition of the matrix X into scores and loadings gives a PCA model that can be written: X = TP’ + E Here, the scores are given by the matrix T and the loadings are given by the matrix P. The variation not explained by the model is given as the residual matrix E. The dimensionality or rank of the model is given by the numbers of score and loading vectors that are calculated. This value is referred to as the number of Principal Components (PCs). A powerful way to visualise the meaning of the PCA model is to plot vectors of scores or vectors of loadings against each other. A score plot from Paper II showing the separation between pine and spruce (figure 5) is a visualisation of two score vectors that represents a plane in the K- dimensional space. Which score vectors to plot can be chosen arbitrarily from the number of components calculated for the PCA model. Usually, it

-21- Multivariate Modelling is most relevant to plot the score vectors that explain most of the variation in the data, i.e. the first and second components. Observations that have similar properties are positioned close to each other while observations of different properties are separated from each other and can normally be found on opposite sides of the plot. The score plots shown in the present thesis have been used for classification of samples. Here, observations that are related to each other often form clusters of objects or so-called groupings. Outliers, i.e. isolated observations that are located far away from the rest, can also be seen in the score plot.

Figure 5. Score plot from 10000 Paper II. Usually, it is 5000 relevant to plot principal components that explain 0 t[2] most of the variation in the -5000 data. It is sometimes

possible to detect groups of -10000 similar objects that are located close to each other. -20000 -10000 0 10000 20000 t[1] SIMCA-P+ 10.5 - 2005-04-11 10:06:31 The corresponding loading plot (figure 6) can then be constructed in order to interpret the score plot and to obtain an understanding of why observations are separated from each other. The loading plot shows variables important for the separation seen in the corresponding score plot. Observations located in a certain area of the score plot are highly correlated to the variables that can be located in the corresponding position in the loading plot. Generally, the most important variables are found far away from the origin of coordinates while variables that are close to origin have very little impact for the plotted principal components. While it is relevant to make two-dimensional loading plots for non-spectral variables, it is sometimes useful to construct a single- component loading plot for spectral data. Figure 7 shows a loading line plot where the values are plotted against their variable numbers, which in this case are the NIR wavelengths of spectra measured on ground wood in Paper IV. This type of plot resembling an original spectrum is used for determining important spectral regions for the plotted principal component.

-22- Multivariate Modelling

0.40 Figure 6. In the loading

0.30 plot, it is possible to detect variables important 0.20 for the separation seen in 0.10 the corresponding score

p[2] plot. Variables located far

0.00 from the origin of

-0.10 coordinates are the most important. -0.20 -0.10 0.00 0.10 0.20 0.30 0.40 0.50 0.60 p[1] SIMCA-P+ 10.5 - 2005-03-21 10:25:51

0.12 Figure 7. The one- 0.1 dimensional loading line 0.08 plot can be relevant to use 0.06 for spectral data in order 0.04 to identify important 0.02

Loading, p1 regions. Figure taken

0 from Paper IV. -0.02

-0.04

-0.06 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm)

PCA was developed by Pearson in 1901 (Pearson) and is now frequently used in studies dealing with multivariate data. PCA has been described in detail by Wold et al. (Wold 1987) and Massart and Heyden have very recently shown how PCA can be used for visualising data tables into plots (Massart and Heyden 2004; Massart and Heyden 2005). In forest applications, PCA has been used for the multivariate characterisation of sample material (Wallbäcks et al. 1991; Lennholm and Iversen 1995; Ferraz et al. 1998).

-23- Multivariate Modelling

5.3 Partial least squares projections to Latent Structures, PLS

LS (Sjöström et al. 1983; Geladi and Kowalski 1986) is a P multivariate method used for finding the relationship between two data matrices. In addition to the X matrix used in PCA, PLS regression also involves a Y matrix or a y vector that can be seen as response variable(s). The use of PLS implies that it is possible to obtain a model that predicts Y from X. In some cases, the y-variable will consist of a dummy variable that gives the class belonging of the corresponding sample. This type of PLS regression is usually referred to as PLS Discriminant Analysis, PLS-DA (Sjöström et al. 1986). As in the case of PCA, PLS reduces the multidimensional space into latent structures in the X block as well as in the Y block where scores for each block are calculated as t and u vectors. The scores are calculated in such a way that the co-variance between X and Y is maximised. For each calculated PLS component, a weight vector, w, is introduced that gives the maximum covariance between the two blocks. PLS regression coefficients, B, can then be derived from the combined weight vectors, W, and from the loadings in X and Y that are denoted as P and C: B = W(P’W)-1C’ The regression coefficients are then used for constructing the prediction model: Y =XB + F It may be useful to construct a number of different plots for the interpretation of PLS models. The one-dimensional weight plot (figure 8) can be used to identify important variables in X that explain much of the variation in y. As in the case of NIR data, this plot taken from Paper I will resemble an original spectrum and the importance of specific spectral regions can be assessed. Another plot of great importance is the observed versus predicted plot (figure 9) for a certain y-variable. Here, observed ISO brightness values for test set samples in Paper IV are plotted against predicted values obtained from the PLS model.

-24- Multivariate Modelling

0.07 Figure 8. In PLS 0.06 modelling, the one- dimensional weight plot 0.05 can be used for finding variables that explain 0.04 much of the variation in

PLS Weights w1* Weights PLS 0.03 the response, Y. Plot taken from Paper I.

0.02

0.01 600 800 1000 1200 1400 1600 1800 2000 2200 Wavelength (nm)

74 Figure 9. An observed vs

73 predicted plot for test set samples in Paper IV can 72 be constructed for the prediction of ISO brightness. 71

Observed 70

67 67 68 69 70 71 72 73 74 Predicted

-25- Multivariate Modelling

5.4 Model Complexity and Validation

ow well a model is estimated according to the variation in the H data is given by the goodness-of-fit, R2, and can be calculated for both X and Y data: R2X = 1 – SS(E) / SS(X) R2Y = 1 – SS(F) / SS(Y) Here, the variances for the matrices E, F, X and Y are given as the Sum of Squares (SS). A R2X or R2Y value close to one unit means that the model explains almost all the variation found in the data.

The determination of the model complexity or rank, which is equal to the number of significant components is a crucial part of PCA and PLS modelling. Too few model components imply that the model is under- fitted and that there still is information left in the data than can be modelled. Too many components, on the other hand, means that the model is over-fitted and variation in the data that is due to noise is modelled. Both these scenarios are unwanted situations and the optimal model rank has to be decided. Cross-validation (Wold 1978) is one method to determine the number of significant model components. The cross-validation procedure implies that a number of samples are omitted during the model calculation and later predicted. The samples are usually divided into a number of cross-validation groups that are left out one at a time so that all samples have been kept out of the model estimation. The Predicted Residual Estimated Sum of Squares (PRESS) can then be calculated for the observations left out of the calculations. PRESS can be used for computing the goodness-of-fit for predicted values, Q2X or Q2Y: Q2X = 1 – PRESS / SS(X) Q2Y = 1 – PRESS / SS(Y) The latter of these two expressions is frequently used for determining both the PLS model complexity and how well the PLS model predicts left out samples.

While the cross-validation gives a rough estimate of how well left out samples are predicted, it is better to use external test sets for the validation of the model. The use of an external test set implies that a number of

-26- Multivariate Modelling samples not included in the model calculation at any time are predicted. The samples used for building the model are called the calibration set. The predicted values can then be compared with the observed or true values of the test set samples. It must be understood that the validation of external test sets should not be used for determining the model complexity. If this is the case, the model is adapted to a particular test set and it is possible that different results could have been obtained if another test set was to be used. Using the predicted test set samples it is possible to calculate the Root Mean Squared Error of Prediction, RMSEP, from the differences between the observed and predicted test set values:

RMSEP = PRESSTS / N where PRESSTS are the predicted residual sum of squares values of the test set samples and N is the number of predicted samples. It is convenient to use RMSEP as a measure of model prediction accuracy due to the fact that it has the same unit as the measured response.

-27-

-28- NIR Spectroscopy

6. NIR Spectroscopy

ear-Infrared spectroscopy or, occasionally, Near-Infrared N Reflectance (NIR) spectroscopy is a technique for measuring transmitted or reflected light of wavelengths between 780 and 2500 nm. The NIR region is consequently positioned between the visual and the infrared regions of the light spectrum. NIR light is absorbed by bonds between hydrogen and the three atoms oxygen, carbon and nitrogen. The absorption is due to overtones or combinations of overtones of the vibrations that occur for the bonds between these atoms. The NIR region was originally discovered by Herschel (McClure 2003) and in the middle of last century Karl Norris lay most of the foundations for how we use NIR spectroscopy today (Butler and Norris 1960; Norris and Rowan 1962). Recently, NIR spectroscopy has frequently been used in research dealing with wood and forest products with many studies focusing on eucalyptus wood species (Schimleck et al. 1996; Michell and Schimleck 1996; Bailleres et al. 2002; Poke et al. 2004). It has also been used for the characterisation of lignin (Yeh et al. 2004) and for evaluating the resin content (Niemz et al.1994). NIR spectroscopy is also commonly used in the food industry for monitoring and quality control (Givens et al. 1997; Buning-Pfaue et al. 1998; McCaig 2002; Buning-Pfaue 2003; Wang et al. 2004; Svensson et al. 2004). The reason for the popularity of the use of NIR spectroscopy in various applications can be deduced from several things. First of all it is a non-destructive method which means that the measured sample is left intact. Secondly, no sample pre-treatment is normally required and, should one be required, it is usually a fairly simple produce. Finally, NIR spectroscopy is a fast and robust method that can be repeated with similar results.

The typical NIR spectrum is comprised of numerous overlapping peaks that arise from the overtone absorption bands. This implies that for most NIR spectra it is impossible to make interpretations based on a single wavelength. The many overlapping peaks in combination with a strong dependency between adjacent wavelengths means that is favourable to

-29- NIR Spectroscopy

analyse a matrix consisting of digitised NIR spectra with multivariate methods. Most studies that use NIR spectroscopy for the characterisation of wood related materials also perform the spectral interpretation by means of multivariate modelling (Borga et al. 1992; Antti et al. 1996; Hauksson et al. 2001; Tsuchikawa and Yamoto 2003; Kelley et al. 2004). Furthermore, NIR spectroscopy is used for determining moisture content (Tsuchikawa et al. 1996; Lestander and Geladi 2003; Blazquez et al. 2004, Nystrom and Dahlqvist 2004) as some of strongest absorption bands in the NIR region are due to bonds in water. The strong correlation between NIR light and water content makes the spectra sensitive to changes in temperature due to wavelength shifts of the hydroxyl absorption bands (Thygesena and Lundqvist 2000; Thygesenb and Lundqvist 2000) but other parts of the spectra are affected as well (Blanco and Valdes 2004). Figure 10 shows NIR spectra of ground wood samples in Paper IV, where the large absorptions around 1450 and 1950 nm are due to the sample water content. For more aspects regarding NIR theory and applications, the reader is referred to the literature (Osborne et al. 1993; So et al. 2004; McClure 2003).

0.8 Figure 10. The 0.7 typical NIR spectra

0.6 consist of broad peaks that often are 0.5 heavily influenced by the water content of 0.4 A the measured samples 0.3 which gives strong

0.2 absorptions around 1450 and 1950 nm. 0.1 0

-0.1 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength nm

-30- NIR Spectroscopy

For Paper I & IV, NIR spectra have been recorded from ground wood and from wood and bark surface. The measurements have been done with a NIR Systems 6500 spectrophotometer (figure 11). This instrument uses a Si based detector for the wavelengths between 400nm to 1098nm and PbS for the wavelengths between 1100 to 2500. Hence, it is possible to record spectra from the visual region with this instrument. The ground wood material has been packed into spinning cups (figure 12) that are inserted in the NIR instrument. The surface of the wood and bark has been measured with a fibre-optic probe (figure 12), which has been held against the analysed material.

Figure 11. The NIR Figure 12. The equipment used with the spectrophotometer, NIR spectrophotometer. The fibre-optic probe Systems 6500, which was to the left and the spinning cup to the used for the measurements right. described in the present thesis.

Both these types of recordings measure the diffuse reflectance (Osborne et al.1993) from the sample. Here, the light is transmitted through the first sample layer where it undergoes absorption. The attenuated remaining light will then be diffused by reflections from other layers of the sample or it can be further transmitted and absorbed. For most samples the light will more or less be subject to scattering, i.e. it will be dispersed in all possible directions because the wavelength is considerably narrower than the size of the particles that the sample is comprised of. This will have a tangible influence on the recorded spectra where the scattering may cause

-31- NIR Spectroscopy additive as well as multiplicative effects. This can mainly be seen as a shift in baseline or slope. Often, an inconsistent sample treatment can cause large variations in the way in which the light is scattered. Figure 13 shows a score plot for the first two components of a PCA model calculated from NIR spectra of ground wood recorded by two different people using the spinning cup (unpublished results). It is evident that the largest variation in the data is caused by the person performing the measurements and, in particular, how that person presented the sample material in the spinning cup.

1.00

0.50

t[2] 0.00

-0.50

-1.00 -10123 t[1] Figure 13. The score plot for the first twoSIMCA-P+ components 10.5 - 2005-02-28 13:48:19of a PCA model calculated from NIR spectra. Two persons recorded the measurements, which form two classes in the plot.

Light scattering is the reason why it is beneficial to perform data pre- treatment on NIR spectra prior to multivariate modelling. Multiplicative Scatter Correction (Geladi et al. 1985), MSC, is a pre-treatment used for removing the additive and multiplicative effects found in spectral data. Without MSC, these effects normally add a few components to a PLS model. A model comprised of raw spectra normally has the same predictive properties as a model calculated from MSC treated spectra, albeit with a higher rank. An extension to the regular MSC, Extended Multiplicative Scatter Correction (Martens el al. 2003), EMSC, has been proposed for samples containing pure ingredients. Orthogonal Signal

-32- NIR Spectroscopy

Correction (Wold et al. 1998), OSC, is another pre-treatment method used for filtering spectral and other kinds of data. The OSC is used in combination with PLS, where it removes variation in the X data that is orthogonal to the response variable(s) y (Y). Orthogonal PLS (Trygg and Wold 2002), O-PLS, is a PLS method with an integrated OSC filter. It divides the X data into orthogonal and non-orthogonal parts. The use of either OSC or O-PLS implies that a single component PLS model can be fitted between filtered NIR spectra and the response y. However, the total number of withdrawn orthogonal components must be taken into consideration when defining the rank of the model. Figure 14 (from Paper I) shows raw NIR spectra, the corresponding MSC treated spectra and the corresponding OSC treated spectra.

1.2 0.9

1 0.8

0.7 0.8 0.6 0.6 log 1/R log 1/R 0.5 0.4 0.4 0.2 0.3 0 0.2 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm) Wavelength (nm)

Figure 14. Raw NIR spectra 1

(above) and the corresponding 0.9 MSC (above right) and OSC 0.8 treated spectra (right). (Paper I) 0.7 0.6

log 1/R 0.5 0.4 0.3

0.2 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm)

-33-

-34- Modelling of Multivariate Images

7. Modelling of Multivariate Images

he digital representation of an image is in terms of rows and T columns of pixels. The pixel properties are given by the intensities of the colour channels or wavelengths that are recorded for the image. Thus, the image representation is a three-dimensional matrix consisting of rows, columns and colour channels (figure 15). Greyscale images only have a single colour channel whereas common RGB images have three (red, green and blue). Other types of images can have even more colours, e.g. NIR cameras (Geladi et al. 2004) can generate up to 80 different colour channels or wavelengths. The number of intensities or levels that is possible for each colour channel is defined by the channel bit depth. For normal RGB images this value is often set to 8 bit, which equals 28 = 256 different intensities. The total number of colours for such an image is 224 = 16777216.

Figure 15. The digital representation of an image consists of rows (N) and columns of pixels (K). Each pixel is then characterised by the intensities (I) in the recorded colour

channels.

N K

The three-dimensional data representing an image contains huge amounts of variation that is preferably analysed by multivariate modelling or by a combination of multivariate modelling and other techniques. Multivariate Image Analysis, MIA, introduced in the late 1980’s and early 1990’s, used PCA for decomposition of digitised images. These studies (Esbensen and Geladi 1989; Geladi 1989; Geladia et al. 1992; Geladib et al. 1992;

-35- Modelling of Multivariate Images

Lindgren and Geladi 1992; Vanespen et al. 1992) compressed images into scores and loadings that visualised the image variation. Multivariate Image Regression (Esbensen et al. 1992; Grahn and Saaf 1992), MIR, was established as a technique where PLS is used for image decomposition according to a y-variable. At this time, MIA and MIR unfolded the image rows and columns to a pixel array (figure 16) that had the number of colour channels as variables.

Figure 16. The original multivariate image analysis unfolded the analysed image into a pixel array (N*K), which had the number of colour channels as variables

(I).

I N K

N*K pixels

Performing PCA or PLS directly on image data is convenient when one image at a time is analysed. However, when numerous images are compared it is often better to use some kind of feature extraction method that collects information from the image data and apply either PCA or PLS. There are methods that analyse the spectral image information, i.e. image colour, and there are also methods that extract spatial information, i.e. the image texture. The image texture can be defined as the pixel variation within the image and where this variation can be located in the image. Here, the local pixel variation will be denoted as frequency information while the location will be denoted as space information. A

-36- Modelling of Multivariate Images review covering most techniques that are being used for texture feature extraction has been written recently (Bharati et al. 2004).

Analysing the spectral information of an image is traditionally done with histograms (Gonzalez and Woods 2002). The histogram construction implies that the pixels of each colour channel are binned according to their intensities. Thus, the number of bins equals 256 for images having 8 bit colour per channel. It has been shown that RGB histograms can be joined with other image features such as HSI (hue, saturation and intensity) values for the analysis of food colour (Antonelli et al. 2004).

The calculation of grey level co-occurrence matrices, GLCM, is a statistical method that can be employed for feature extraction of an image. The co-occurrence matrix (Haralick 1973) is a two-dimensional table that uses pixel intensities for its X and Y axes. Each point in the matrix gives the probability that there are two pixels with the intensities given by the two axes at a certain distance. The distance is in pixels and is an arbitrary number that can be set according to the pixel variation of the analysed image. It is also necessary to give the direction for the distance. The distance can be set to horizontal, vertical or diagonal directions. Preferably, the direction should be set so that it corresponds to the structural features of the image texture. Compared with other methods, it has been shown that co-occurrence matrices can be used for efficient texture feature extraction (Ohanian and Dubes 1992). More studies have shown (Basset et al. 2000; Foucerot et al. 2004) that co-occurrence matrices can be used for analysing image texture.

Furthermore, image transformations can be used for extracting features from images. The two-dimensional discrete Fourier Transform (2D DFT) can be used for capturing the frequency information of the image texture (Geladi 1992). This implies that information about space is lost as the 2D DFT transforms image data into the frequency domain.

On the other hand, the two-dimensional Discrete Wavelet transform (2D DWT) is a transformation that can be applied to image data for capturing both frequency and space information. This transformation uses small waves, or so-called wavelets for the image decomposition (Daubechies

-37- Modelling of Multivariate Images

1992). The wavelets are used for analysing images at many different resolutions, denoted as scales. The use of the wavelet transform on different kinds of data may serve different purposes. It can be done in order to perform noise-reduction (Mallat and Hwang 1992; Deng 1996; Cai et al. 2001; Cho and Bui 2005) and for data compression (Devore et al. 1992; Luo et al. 1997; Grgic et al. 2001; Trygg et al. 2001). Its use for capturing texture features has been proven in numerous studies (Laine and Fan 1993; Henkereed and Cheng 1993; Unser 1995; Van de Woewer et al. 1999; Bashar et al. 2003). The reader is referred to a textbook (Gonzalez and Woods 2002) for more information and theory about wavelet transformations.

As mentioned earlier, the original MIA and the MIR techniques analysed one image at a time where each pixel represented an observation or sample in the constructed multivariate models. However, for classification of many images at a time, where each image is considered as an observation, it is more convenient to extract features from the images prior to calculating multivariate models. Recently, a number of studies that calculates multivariate models on extracted image data have been shown (Kvaal et al. 1998; Yu and MacGregor 2003, Antonelli et al 2004). This approach has also been used for classification of spruce and pine bark in Paper II that will be discussed in the next chapter. The extraction of textural features by calculating co-occurrence matrices was an optional step in the image processing prior to the 2D DWT and 2D DFT transformations. 2D DWT has been used for capturing and compressing the image information. 2D DFT has then been applied to each wavelet scale in order to remove space information. For the image classification conducted in this thesis only frequency information was considered important as the location of small sample variations within the image are not related to the wood species of either spruce or pine. It should be pointed out that the use of 2D DWT was paramount for the feature extraction in one aspect, namely that it was possible to summarise most of

-38- Modelling of Multivariate Images the image information in one Figure 17. The image analysis in vector consisting of wavelet Paper III uses extraction coefficients obtained from the techniques that compress the processed wavelet scales. image information into one Further, 2D DFT was not applied vector for each image. The to the wavelet scales derived vectors of many images are then from the co-occurrence matrices. assembled as a matrix that can The information found in a co- be analysed by means of occurrence matrix is related only multivariate data analysis. to the image frequency information due to the fact that compared intensities disregard the exact location of image pixels. The extracted features Digitalisation into from each image were thus given colour channels as one vector regardless of origin (figure 17). The vectors were then assembled as matrices comprised of spruce and pine samples as observations and histograms or wavelet Feature extraction coefficients as variables. PCA and PLS were then utilised for constructing multivariate models that was used for the classification of pine and spruce.

PLS PCA

2,00 10000 1,80 1,60 0 t[2] 1,40

YVarPS(y) -10000 y observed y 1,20 -20000 1,00

0,90 1,00 1,10 1,20 1,30 1,40 1,50 1,60 1,70 1,80 1,90 2,00 -30000 -20000 -10000 0 10000 20000 YPredPS[6](y)y predicted t[1]

SIMCA-P+ 10.5 - 2005-02-08 15:56:15 SIMCA-P+ 10.5 - 2005-02-08 11:02:53

-39-

-40- Roundwood Species Classification

8. Roundwood Species Classification

he two softwood species Norway spruce (Picea Abies) and T Scots pine (Pinus Sylvestris) represent, according to the Swedish sawmill inventory, 98% of the roundwood transported to the Swedish sawmills. Some mills have concentrated the production using only one species at a time but the majority of sawmills is, in fact, processing both species. When the two species are sawn at the same time, the incoming roundwood needs to be classified at some point in the production chain. This is normally done by manual assessment of the logs prior to the scaling operation. At the moment, no automatic methods are used for the species classification of roundwood. There are, however, techniques for automatic grading of logs such as gamma-ray (Hagman 1993) and X-ray (Oja et al. 2003). There are also techniques for the classification of sawn wood including FT-NIR spectroscopy (Brunner et al. 1996) and ultrasonic signalling (Jordan et al. 1998). In this thesis, two methods that could be used to automatically assess the softwood wood species are presented. These two methods, NIR spectroscopy and multivariate image analysis, have been used for analysing the bark of the samples. However, the wood surfaces of the samples have also been analysed as the tree felling and transportation sometimes involves that the bark is partly peeled off.

8.1 Roundwood Species Classification using NIR spectroscopy, Paper I

he data set used in this study comprised of 53 wood discs that T were collected from a thinning. NIR spectroscopy was used to record spectra of bark and wood at wavelengths between 400 to 2500 nm. Every second wavelength was recorded and the spectra were assembled as two matrices sized 53x1050 and 48x1050. The score plot for the first

-41- Roundwood Species Classification three components (figure 18) of the PCA model calculated for the spectra recorded from bark showed that the spruce objects form a distinct group in the middle while the pine samples are divided into two smaller groups.

Figure 18. The score plot for the first three components of a PCA model calculated from NIR spectra of pine and spruce bark. The spruce samples form a group in the middle while the pine

sample are scattered on both sides.

The reason why the pine samples were divided into two groups was due to the fact that the bark is of a smoother nature at the top of the stem while it is rougher at the bottom. The PLS model between the bark spectra and the wood species belonging as y-variable showed that it was possible to correctly classify all excluded test set samples. The corresponding PLS model based on wood spectra was, with a few wrongly predicted samples, not as accurate. This type of model could, however, prove to be important if the classified sample had the barked peeled off for some reason. An analysis of the PLS weights for the classification of the bark samples showed that the most important wavelength region was around 800nm, which corresponds to the upper visual part and the lower near-infrared part of the spectrum. It was also shown that the wavelengths around 2000 nm were important. Reduced PLS models, using either the lower part of the spectrum or the upper part of the spectrum only, proved to give as efficient predictions as the full model. This may allow for a simpler instrumentation setup for an automated classification system using NIR spectroscopy.

-42- Roundwood Species Classification

8.2 Roundwood Species Classification using Multivariate Image Analysis, Paper II.

he sample material for this study consisted of 146 discs from T spruce and pine logs collected from a clearing. RGB images that contained bark, wood and a combination of bark and wood were recorded from the surface of each disc. An overview of the study can be seen in figure 19. Extraction of features as discussed in the chapter five was applied to the images. Extracted features included: 1. Coefficients from Fourier transformed wavelet compressed images 2. Coefficients from wavelet compressed GLCM data

Overall, the best classification was obtained by using wavelet compression on GLCM data. The score plot for the first two components of a PCA model derived from GLCM of images containing bark (figure 5 in Chapter 5) shows that there is an obvious separation between the two classes. Only one out of 48 test set samples was wrongly predicted in the PLS model. The classification using images containing wood only was not as successful, with 73% prediction accuracy. In order to simulate a scenario that describes a real situation, the images containing both bark and wood were predicted. All three image types were divided into smaller sub-images that were used for building a PCA classification model. It can be seen in the score plot for the first two components (figure 20) that the bark is easily separable from the wood.

15000 Figure 20. The score plot 10000 for the first two components of a PCA model 5000 derived from wood and t[2] 0 bark images. The black objects are wood samples, -5000 the grey are bark samples

-10000 and the light grey are

-20000 -10000 0 10000 20000 images containing both t[1] bark and wood. SIMCA-P+ 10.5 - 2005-04-12 19:08:15

-43- Roundwood Species Classification

67 pine samples 79 spruce samples

Spruce bark Spruce bark+wood Spruce wood Pine bark Pine bark+wood Pine wood

Figure 19. An overview 1. 2. of the wood species GLCM classification with MIA. The percentages correctly classified observations are 2D DWT 2D DWT given next to the score plots. 2D DFT

Images represented by vectors

15kCoefficients 15kCoefficients 1 1 Images Images Data matrices 146 146

Test set samples

20000 10000

10000 5000

0 0 t[2] t[2]

-5000 -10000

-10000 -20000

-20000 0 20000 40000 -20000 -10000 0 10000 20000

t[1] t[1]

SIMCA-P+ 10.5 - 2005-05-13 12:00:54 SIMCA-P+ 10.5 - 2005-05-16 09:07:22 90% 98%

20000 10000

10000 5000

0 0 t[3] t[2] -10000 -5000 -20000 -10000 -30000

-40000 -20000 0 20000 40000 -10000 0 10000

t[2] t[1]

SIMCA-P+ 10.5 - 2005-05-13 12:05:17 SIMCA-P+ 10.5 - 2005-05-16 09:05:34 81% 85%

10000

10000 5000

0 t[2] t[2] 0 -5000

-10000 -10000

-20000 0 20000 -20000 -10000 0 10000 20000

t[1] t[1]

SIMCA-P+ 10.5 - 2005-05-16 09:38:36 SIMCA-P+ 10.5 - 2005-05-16 09:02:12 67% 73%

-44- Roundwood Species Classification

The sub-images originating from the images containing both bark and wood are spread among the two classes. These images were then used as an external test set but only those images that were classified as bark were predicted. The pure bark images was here used as calibration set for the PLS model that managed to correctly classify 92% of the samples in the test set.

The pure wavelet classification could predict the unknown test set comprised of bark images with an accuracy of around 90%. The weight plot for the first two components (figure 21) of the PLS model calculated from the wavelet compressed images shows that only the coefficients accounting for the mean colour of the samples contributed to the separation between the two wood species. Evidently, it is difficult to construct prediction models that rely on structural information without extracting features prior to the modelling. Although the wavelet compression in combination with the Fourier transformation can be seen as a feature extraction method in itself, this approach merely summarises all the information that can be found in the images.

0.50

0.40

0.30

w*[2] 0.20

0.10 0.00

-0.10 -0.60 -0.40 -0.20 0.00 0.20 0.40 0.60 w*[1] Figure 21. The loading plot for theSIMCA-P+ first 10.5two - 2005-02-08 components 11:34:52 of a PCA model calculated from wavelet compressed images. The only important wavelet coefficients (marked with arrows) are those that account for the mean colour information.

-45- Roundwood Species Classification

8.3 Comparison between the two methods

he results for the wood species classification show that the T method using NIR spectroscopy has similar or even better prediction accuracy than the method using multivariate image analysis. There are, however, a few circumstances that could favour the use of image analysis instead of NIR spectroscopy: 1. Instrumentation setup. The NIR instrumentation should be fairly simple. It has been shown that an NIR instrument can be used online for monitoring incoming sample material (Feldoff et al. 1997; Axrup et al. 2000; Anderson and Walker 2003; Kulcke et al. 2003). On the other hand, the camera setup could be even simpler. Using co-occurrence feature extraction, the present study has shown that low resolution greyscale images can successfully predict bark images. 2. Robustness of the instrumentation. Conditional variations such as distance to the measured sample material, variable light conditions and temperature affect the NIR instrumentation. While these problems may be solved, they definitely need to be taken into account. A camera setup is rather insensitive to the sample distance provided that a sufficient depth of field is being used. The camera should also be insensitive to changes in temperature. While changes in ambient light might cause problems, this can easily be solved by putting some kind of reference in the line of sight of the camera. 3. Future model improvements. The PLS prediction models based on NIR spectra have been constructed under laboratory conditions. While the number of included calibration set objects can be increased in order to cover a larger variation in sample material, real conditions might lessen the predictive ability. An advantage with the NIR classification is that the actual prediction models are very simple. This is also a drawback as it probably means that the models are optimal regarding model complexity and data pre-treatment. In other words, there is little room for model improvements. The multivariate image analysis on the other hand, uses more complex prediction models. These models should be largely uninfluenced by industrial conditions due to the robustness of the instrumentation. There should also be substantial room for improvements, as much of the image feature extraction is still unexplored.

-46- Roundwood Species Classification

For both methods, stepwise classification could be possible as they are sufficiently fast to record data from more than one point of the incoming log. The predictions made in the present thesis have used a cut-off (e.g. 0.5 or 1.5) that is exactly between the values giving the true class belonging (e.g.1 or 2). It might, however, be feasible to use cut-off values that are closer to the true values, where doubtful samples could be predicted once more. This next step could be another data recording or in case image analysis is being used, just another feature extraction method. It might even be possible to use a combination of both techniques in an instrumentation setup used for automatic wood species classification.

-47-

-48- Wood Storage and its Relation to the Brightness of Bleached TMP

9. Wood Storage and its Relation to the Brightness of Bleached TMP

ach year enormous quantities of conifer trees are felled in E Sweden. In year 2000, over 33 million m3 of wood raw material was transported to the saw mills alone according to the Swedish sawmill inventory. The total wood consumption where pulp mills are included is over 80 million m3. The logistics system for felled roundwood involves storage at different locations. Usually, roundwood is stored at the roadside for some time before the logs are transported to either a pulp or saw mill. In order to ensure a shortage-free production a certain wood supply has to be present at the mills. This type of storage normally takes place during a longer period of time compared to the roadside storage. As mentioned in the introduction, the storage often has undesirable effects on the quality of wood. This is a matter of great interest and has recently been dealt with in two doctoral theses (Liukko 1997, Persson 2003).

Especially during summertime, the storage brings unwanted effects to the quality of the stored wood and this is due to several reasons. In order to avoid damage caused by drying, roundwood transported to the saw mills is usually sprinkled with water (Elowson and Liukko 1995; Liukko 1997; Persson 2001; Persson and Elowson 2001). Apart from apparent drying damages such as checking and debarking, the drying of wood makes it vulnerable to fungi, causing blue stain and to beetle attacks (Liukko 1995, Persson 2001). Sprinkling or wet storage can, however, cause unwelcome effects with the respect to brightness of wood (Lorås 1974; Persson and Elowson 2001, Persson et al. 2002). The risk of bacterial damage and fungi causing brown rot increases with a higher moisture content (Liukku 1997). One of the major drawbacks of wet storage or sprinkling is the discolouration of wood that occurs when organic compounds, mostly tannins, migrate from the bark into the sapwood (Lorås 1974; Wilhelmsen 1968; Persson et al. 2002). This so-called bark staining decreases the bleachability of pulp produced from the stored wood. Moreover, it is

-49- Wood Storage and its Relation to the Brightness of Bleached TMP generally considered that metal ions of Mn, Fe and Ca can migrate into the wood during wet storage. A high metal ion content in the wood may increase the rate of the decomposition of peroxide which is one of the chemicals most often used for bleaching of thermo mechanical pulp. However, a recent study (Persson and Lindeberg 2001) has shown that while there was a certain accumulation of metal ions during sprinkled storage, this increase was marginal and could not be related to large brightness reductions for bleached thermo mechanical pulp.

The bleaching of thermo mechanical pulp is done in order to remove coloured compounds of the lignin (Dence and Reeve 1995). These compounds, sometimes called chromophores, react with the used bleaching agent which leads to a brightness increase of the bleached mechanical pulp. The chromophore groups mainly consist of carbonyl, ethylenic and phenolic rings that form conjugate systems and thus give rise to wood or pulp colouration (Dence and Reeve 1995). Two common agents used in mechanical pulping are hydrogen peroxide, H2O2, and sodium dithionite, Na2SO4.

Hydrogen peroxide is used under alkaline conditions, which leads to the formation of the hydroperoxide ion: − − HOOH + HO ←⎯→ HOO + H 2O The hydroperoxide ion is a very strong nucleophile and reacts with lignin mostly in an oxidative manner. The known reactions of the chromophore eliminations are described in detail by Dence and Reeve (Dence and Reeve 1995). Furthermore, it is well known that hydrogen peroxide decomposes into water and oxygen:

2HOOH ⎯⎯→ 2H 2O + O2 The decomposition of hydrogen peroxide is catalysed by metal ions such as Fe3+ and Mn2+. Various aspects of peroxide bleaching and its stability in presence of metal ions have been covered in many studies (Abbot and Hobbs 1992; Gierer and Jansbo1993; Brown and Abbot 1995). Owing to the catalytic decomposition of peroxide it is conventional that stabilising agents such as EDTA and DTPA are used as metal chelants for forming complexes with metal ions (Brown and Abbot 1995). Another reason to treat the pulp with chelants is due to the formation of coloured complexes

-50- Wood Storage and its Relation to the Brightness of Bleached TMP between metal ions and structures in the lignin (Ghosh and Ni 1998, Friman et al. 2004).

Sodium dithionite is contrary to peroxide a reductive agent that during its reaction with lignin converts into the bisulphite ion (Dence and Reeve 1995): 2− − + − S 2O4 + 2H 2O ⎯⎯→ 2HSO3 + 2H + 2e Sodium dithionite or sometimes hydrosulphite will oxidise in the presence of oxygen and form a mixture of various sulphur compounds. The decomposition of dithionite has been investigated in a number of studies (Lister and Garvie 1959; Rinker 1965, Kilroy 1980). The efficiency of dithionite bleaching is affected by the pH value and this has also been a subject for some interest (Tredway 1979; Svensson 1998).

Moreover, it is a well-known fact the brightness of mechanical pulp can also be degraded by ultraviolet or visible radiation that is absorbed by the chromophoric conjugated systems of lignin (Gellerstedt and Petterson 1978; Agnemo et al. 1991). The effect of intensity has been explored in a relative recent study (Andrady 1999) and the use of stabilising agents to retard this so-called photo-yellowing has also been investigated (Cook et al. 1996; Cook and Ragauskas 1997). In brief, there are many factors that might influence the final brightness of bleached mechanical pulp. The brightness is normally given in terms of ISO brightness which is a well- known standard in the Pulp and Paper industry. An elaboration of the meaning of ISO brightness can be obtained from a couple of papers published in the last decade (Bristow 1994, Bristow 1999).

9.1 The Effect of Designed Wood storage on the Brightness of Bleached TMP, Paper III

he wood storage described in Paper III was a designed study T with five factors that controlled the storage conditions. Sixty logs of Norway Spruce were collected in a forest district in the middle of Sweden and transported to climate chambers for storage. The five factors

-51- Wood Storage and its Relation to the Brightness of Bleached TMP light, temperature, watering, tree growth and level of debarking were varied according to a D-optimal design that gave different conditions for each of the 15 experiments. Samples (wood chunks) were collected in the beginning, in the middle, and at the end of the storage. These chunks were completely debarked and refined into 45 samples of thermo mechanical pulp. The TMP was bleached with either hydrogen peroxide or sodium dithionite so that the final sample material consisted of 45 unbleached, 45 peroxide bleached and 45 dithionite bleached pulps. The ISO brightness was then measured for each of the bleached pulps. The design table, X, was then fitted to the measured ISO brightness values, Y, using three PLS models which R2Y and Q2Y values are given in figure 22.

2 1 Figure 22. The R Y and 2 0,9 Q Y values of the models 0,8 calculated for 0,7 unbleached, dithionite 0,6 R2Y 0,5 bleached and peroxide Q2Y 0,4 bleached TMP. The

0,3 model for the peroxide

0,2 bleached TMP has the 0,1 highest degree of 0 Unbleached TMP Dithionite Bleached Peroxide Bleached explanation among the TMP TMP three models.

The effect of time was added as an extra factor and the Y values were expressed as the loss in ISO brightness compared to brightness of the fresh samples. The PLS models could then be used for identification of the most important factors affecting the brightness and thus bleachability of unbleached, peroxide bleached and dithionite bleached TMP. An overview of the storage experiments, which includes Paper IV, can be seen in figure 23. The evaluation of the results showed that the brightness decrease for the unbleached pulp was around three ISO brightness units for three replicates and this is in line with what was reported in an earlier study (Persson et al. 2002). These replicates had been subjected to a relatively high level of watering so the degree of bark staining was substantial. The coefficient plot derived from the model calculated for all of the unbleached pulp samples (figure 24) showed that the most signifi-

-52- Wood Storage and its Relation to the Brightness of Bleached TMP

Some logs were freed from outer bark

Stand with fast growing spruce

Wood samples consisting of four logs each Stand with slow growing spruce

Sampling Storage in climate chamber Bleached pulp TMP Dithionite refining Chipping

Peroxide

NIR measurements

0.8 0.7 ISO brightness measurements 0.6 0.5 0.4 A 0.3 0.2 0.1 0 -0.1 400 600 800 10001200140016001800200022002400 Wavelength nm NIR spectra ISO brightnessvalues samples Storage samples NIR spectra Variables Bark Temperature Light Tree Growth Tree Watering

wavelengths 81

80.5 PLS modelling 80

Prediction of dithionite 79.5

bleached pulp 79

Observed 78.5 PLS XY 78 77.5 77 77 77.5 78 78.5 79 79.5 80 80.5 81 Predicted Prediction of peroxide RMSEP = 0.96 bleached pulp

Unbleached TMP 74

1,50 1,00 0,50 Coefficients for storage 73 0,00

-0,50 factors Lig Lig wat Tim Tim 72 temp temp wat*Lig wat*Lig wat*Tim wat*Tim Dithionite bleached TMP 1,50 71

1,00 0,50 Observed 70 0,00

-0,50 Lig wat Tim wat*Lig wat*Tim 69 Peroxide bleached TMP 0,70

0,60

0,50 68

0,40

0,30

0,20

0,10 67

0,00 67 68 69 70 71 72 73 74 wat temp Predicted RMSEP = 0.74 Figure 23. An overview of the two wood storage studies of the thesis. Ultimately, PLS models that give the importance of the storage factors and that predict the ISO brightness of bleached TMP can be obtained.

-53- Wood Storage and its Relation to the Brightness of Bleached TMP cant term was the interaction between watering and light, followed by the terms for temperature and time. The terms for watering and the interaction between watering and time were close to being significant according to 95% confidence level. The term for light was insignificant but, in conjunction with watering, it is clear that these two terms account for a considerable brightness loss. The size of each of the coefficients can be translated into the brightness loss that occurs when the particular factor is varied from centre to high level. It is noticeable that the level of debarking and the tree growth had no to little effect on the brightness. It should be stated though; that a high debarking level corresponded to a removal of only outer bark. Moreover, the model fit of the PLS model indicated that much variation was not modelled. This is most likely due to uncertainties in the y- variable.

Unbleached TMP Figure 24. The 1,50 coefficients of the models 1,00 calculated from the three types of TMP samples. 0,50 The unbleached and 0,00 dithionite bleached pulp -0,50 Lig

wat brightness is heavily Tim temp wat*Lig wat*Tim dependent on the

Dithionite bleached TMP interaction between 1,50 watering and light. The

1,00 brightness of peroxide bleached pulp is on the 0,50 other hand dependent on 0,00 the effects of watering and -0,50 temperature. Lig wat Tim wat*Lig wat*Tim Peroxide bleached TMP 0,70 0,60 0,50 0,40

0,30 0,20 0,10

0,00 wat temp

-54- Wood Storage and its Relation to the Brightness of Bleached TMP

With the exception of temperature, the coefficient plot for the dithionite bleached pulps includes all the terms used for modelling the brightness loss of the unbleached pulps. The brightness of dithionite bleached TMP is, therefore, dependent on all these terms. Considering the fact that the coefficients are comparable in size to the ones obtained from the unbleached samples, this means that dithionite bleaching is unable to remove brightness losses caused by these terms. Dithionite bleaching can, however, be used for removing effects introduced by a high temperature.

As seen in the coefficient plot for the peroxide bleached samples, only the terms for watering and temperature are significant. This implies that peroxide bleaching can be used for removing the brightness loss caused by the interaction of light and watering. It also means that bleachability of TMP did not decrease with increased storage time. Hence, most of the brightness reversion for the bleached TMP occurred during the first six weeks. Nevertheless, peroxide bleaching seems to be sensitive to brightness reversion introduced by watering and temperature.

As mentioned earlier, the risks of wood storage regarding the bleachability of TMP are highly pronounced during summer-time. This has also been confirmed by the present study where watering, light and temperature cause a brightness reversion of bleached TMP. The most significant effect for the brightness of unbleached and dithionite bleached was the interaction between watering and light. Peroxide, on the other hand, is totally unaffected by the brightness loss caused by this interaction. The interaction between watering and light could be related to an increased activity of micro-organisms, which is affected by the accessibility of light and water. Furthermore, both peroxide and dithionite bleaching seem sensitive to the brightness reversion caused by watering, even though the coefficient for dithionite is not significant according to 95% confidence level. As shown in earlier studies (Lorås 1974; Persson et al. 2002) this effect is due to bark staining. Finally, it is noteworthy that dithionite can be used for removing the effect caused by temperature in contrary to peroxide.

-55- Wood Storage and its Relation to the Brightness of Bleached TMP

9.2 Prediction of Bleached TMP Brightness using NIR Spectroscopy, Paper IV

his study is based on wood material collected from the logs T during the wood storage explained in Paper III. These samples were taken every second week and consisted of wood discs with a thickness of approximately 5 cm. NIR spectroscopy was then used for recording spectra from four different kinds of material: 1. Peripheral surface of each disc 2. Ground wood 3. Dried ground wood 4. Unbleached TMP The measurements also included the visual region of the spectra. PLS was then used for calculating prediction models between the spectral data and the ISO brightness of the bleached thermo mechanical pulp. The measurements on the unbleached TMP were performed just for the sake of comparison.

It was seen that the brightness of dithionite bleached TMP could be predicted from the peripheral wood surface (figure 25), which was not the case for peroxide bleached TMP.

74 Figure 25. The observed

73 vs predicted plot made for test set samples predicted 72 from NIR spectra recorded from wood 71 surface.

Observed 70

67 67 68 69 70 71 72 73 74 Predicted

-56- Wood Storage and its Relation to the Brightness of Bleached TMP

This consolidates the findings made in the previous paper where the strong interaction effect between watering and light played a dominating part in the bleachability of TMP when using dithionite as bleaching agent. This interaction should mostly affect the outer parts of the sapwood that is close to the bark. The models made for the ground wood material could be used for predicting the ISO brightness of both dithionite and peroxide bleached TMP. It was shown, however, that the predictive ability of the models that originated from the near-infrared region of the spectra was limited without drying the samples prior the spectral measurements. The analysis of PLS weights showed that wavelengths corresponding to the near-infrared region was highly influenced by absorptions related to bonds of water (figure 26). Moreover, the weight plots showed that the most important parts of the spectra corresponded to the visual region. It was also seen that much of the variation of the data was orthogonal to the predicted ISO brightness and could therefore be removed by means of OSC prior to the computing of PLS prediction models.

0.12 Figure 26. The analysis of 0.1 PLS weights showed that 0.08 the modelling was heavily 0.06 influenced by the sample 0.04 water content. 0.02 PLS Weights Weights w PLS 0

-0.02

-0.04

-0.06 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm)

The modelling suggests that is should be possible to build prediction models from NIR spectra of stored wood that can be used for predicting a rough estimation of the ISO brightness achieved after dithionite bleaching. If a model is to be used for predicting the brightness obtained after a subsequent peroxide stage, the wood has to be ground prior the spectral recordings.

-57-

-58- Concluding Remarks

10. Concluding Remarks

he studies in this thesis have shown that the use of a T multivariate approach is a prerequisite for finding different properties of roundwood. The two methods used for the wood species classification generated large quantities of data that were efficiently handled by PCA and PLS models. The study using multivariate image analysis demonstrated that it was also necessary to extract certain image features prior to the multivariate modelling. The two classification studies showed that the wood species could be predicted from the bark of the analysed log. The measurements were, however, carried out in a laboratory environment which means that other results could have been obtained if an industrial environment was used.

The use of experimental design in the brightness study, revealed that it was possible to identify factors that negatively affected the bleachability of TMP. This made it possible to identify the large interaction between watering and light that affected the bleaching of TMP using dithionite. It is virtually impossible to detect such interactions without employing a designed experiment. The use of experimental design also indicated that the samples used for constructing PLS prediction models for the ISO brightness of bleached TMP were as diverse as possible. Picking samples randomly often produces narrow data sets that cannot be used in a more general context.

-59-

-60- References

11. References

Abbot, J. and G.C. Hobbs (1992). “The Influence of Metal-Ions on 2- Stage Peroxide Bleaching of Radiata Pine Tmp.” Journal of Pulp and Paper Science 18(2): J67-J70. Agnemo, R., R.C. Francis, T.C. Alexander and C.W. Dence (1991). ”Studies on the Mechanism of the Photoyellowing of Bleached Mechanical and Chemimechanical Pulps .3. The Role of Hydroxyl Radicals.” Holzforschung 45: 101-108. Anderson, N.M. and P.N. Walker (2003). ”Measuring fat content of ground beef stream using on-line visible/NIR spectroscopy.” Transactions of the Asae 46(1): 117-124. Andrady, A.L. and A. Torikai (1999). ”Photoyellowing of mechanical pulps - III. Intensity effects and dose-response relationships.” Polymer Degradation and Stability 66(3): 317-322. Antonelli, A., M. Cocchi, P. Fava, G. Foca, C. G. Franchini, D. Manzini and A. Ulrici (2004). ”Automated evaluation of food colour by means of multivariate image analysis coupled to a wavelet-based classification algorithm.” Analytica Chimica Acta 515(1): 3-13. Antti, H., M. Sjöström and L. Wallbacks (1996). ”Multivariate calibration models using NIR spectroscopy on pulp and paper industrial applications.” Journal of Chemometrics 10(5-6): 591-603. Axrup, L., K. Markides and T. Nilsson (2000). ”Using miniature diode array NIR spectrometers for analysing wood chips and bark samples in motion.” Journal of Chemometrics 14(5-6): 561-572. Bailleres, H., F. Davrieus and F.H. Pichavant (2002). ”Near infrared analysis as a tool for rapid screening of some major wood characteristics in a eucalyptus breeding program.” Annals of Forest Science 59(5-6): 479-490. Bashar, M.K., T. Matsumoto and N. Ohnishi (2003). ”Wavelet transform- based locally orderless images for texture segmentation.” Pattern Recognition Letters 24(15): 2633-2650. Basset, O., B. Buquet, S. Abouelkaram, P. Delachartre and J. Culioli (2000). ”Application of texture image analysis for the classification of bovine meat.” Food Chemistry 69(4): 437-445.

-61- References

Bharati, M.H., J.J. Liu and J.F. MacGregor (2004). ”Image texture analysis: methods and comparisons.” Chemometrics and Intelligent Laboratory Systems 72(1): 57-71. Blanco, M. and D. Valdes (2004). ”Influence of temperature on the predictive ability of near infrared spectroscopy models.” Journal of near Infrared Spectroscopy 12(2): 121-126. Blazquez, C., G. Downey, C. O'Donnell, D. O'Callaghan and V. Howard (2004). ”Prediction of moisture, fat and inorganic salts in processed cheese by near infrared reflectance spectroscopy and multivariate data analysis.” Journal of near Infrared Spectroscopy 12(3): 149-157. Borga, P., M. Hamalainen and O. Theander (1992). ”Correlations between near-Infrared Spectra of Wet-Stored Timber and the Storage Time in Relation to the Water-Quality.” Holzforschung 46(4): 299-303. Box, G.E.P and R.D. Draper (1987). Empirical Model-Building and Response Surfaces. John Wiley and Sons. New York Bristow, AB. (1994). “What is ISO Brightness?” Tappi Journal 77(5):174. Bristow, AB. (1999). “ISO Brightness, a more Complete Definition.” Tappi Journal 82(10):54. Brown, D.G. and J.Abbot (1995). “Effects of Metal Ions and Stabilisers on Peroxide Decomposition during Bleaching.” Journal of Wood Chemistry and Technology 15(1):85. Brunner, M., R. Eugster, E. Trenka and L. BerganminStrotz (1996). ”FT- NIR spectroscopy and wood identification.” Holzforschung 50(2): 130-134. Buning-Pfaue, H. (2003). ”Analysis of water in food by near infrared spectroscopy.” Food Chemistry 82(1): 107-115. Buning-Pfaue, H., R. Hartmann, J. Harder, S. Kehraus and C. Urban (1998). ”NIR-spectrometric analysis of food. Methodical development and achievable performance values.” Fresenius Journal of Analytical Chemistry 360(7-8): 832-835. Butler, W.L. and K.H. Norris (1960). “The spectrophotometry of dense light-scattering material.” Archives of Biochemistry and Biophysics 87(1): 31-40. Cai, Z.H., T.H. Cheng, C. Lu and K.R. Subramanian (2001). ”Efficient wavelet-based image denoising algorithm.” Electronics Letters 37(11): 683-685. Cho, D. and T.D. Bui (2005). ”Multivariate statistical modeling for image denoising using wavelet transforms.” Signal Processing-Image Communication 20(1): 77-89.

-62- References

Cook, C.M. and A.J. Ragauskas (1997). ”Brightness reversion of mechanical pulps .8. Investigation of synergistic photostabilization methods for high-yield pulp.” Journal of Photochemistry and Photobiology a-Chemistry 104(1-3): 217-224. Cook, C.M., X. Pan and A. J. Ragauskas (1996). ”Brightness reversion of mechanical pulps .7. Photostabilization studies of thiol additives for lignocellulosic materials.” Journal of Wood Chemistry and Technology 16(3): 327-345. Daubechies, I. (1992). Ten Lectures on Wavelets. SIAM, Philadelphia. Dence, C.W. and D.W. Reeve (1996). Pulp Bleaching; Principles and Practise, Tappi Press, Atlanta. Deng, G. (1996). ”Greyscale image representation using binary sketch image.” Electronics Letters 32(10): 881-882. Devore, R.A., B. Jawerth and B.J. Lucier (1992). ”Image Compression through Wavelet Transform Coding.” Ieee Transactions on Information Theory 38(2): 719-746. Elowsson, T. and K. Liukko (1995). ”How to achieve effective wet storage of pine logs (Pinus sylvestris) with a minimum amount of water.” Forest Products Journal 45(11-12): 36-42. Esbensen, K. and P. Geladi (1989). ”Strategy of Multivariate Image- Analysis (Mia).” Chemometrics and Intelligent Laboratory Systems 7(1-2): 67-86. Esbensen, K.H., P.L. Geladi and H.F. Grahn (1992). ”Strategies for Multivariate Image Regression.” Chemometrics and Intelligent Laboratory Systems 14(1-3): 357-374. Feldhoff, R., D. Wienke, K. Cammann and H. Fuchs (1997). ”On-line post consumer package identification by NIR spectroscopy combined with a FuzzyARTMAP classifier in an industrial environment.” Applied Spectroscopy 51(3): 362-368. Ferraz, A., E. Esposito, R.E. Bruns and N. Duran (1998). ”The use of principal component analysis (PCA) for pattern recognition in Eucalyptus grandis wood biodegradation experiments.” World Journal of Microbiology & Biotechnology 14(4): 487-490. Foucherot, I., P. Gouton, J.C. Devaux and F. Truchetet (2004). ”New methods for analysing colour texture based on the Karhunen- Loeve transform and quantification.” Pattern Recognition 37(8): 1661-1674. Friman, L., H. Hoglund, H.E. Hogberg and R. Agnemo (2004). ”Tannin- iron impregnated thermomechanical pulp - Part II: Bleachability and brightness reversion.” Nordic Pulp & Paper Research Journal 19(4): 525-531.

-63- References

Geladi, P. (1992). ”Some Special Topics in Multivariate Image-Analysis.” Chemometrics and Intelligent Laboratory Systems 14(1-3): 375- 390. Geladi, P. and B.R. Kowalski (1986). ”Partial Least-Squares Regression - a Tutorial.” Analytica Chimica Acta 185: 1-17. Geladi, P., D. MacDougall and H. Martens (1985). “Linearization and scatter-correction for NIR reflectance spectra of meat.” Applied Spectroscopy 39(3): 491-500. Geladi, P., H. Isaksson, L. Lindqvist, S. Wold and K. Esbensen (1989). ”Principal Component Analysis of Multivariate Images.” Chemometrics and Intelligent Laboratory Systems 5(3): 209-220. Geladi, P., J. Burger and T. Lestander (2004). ”Hyperspectral imaging: calibration problems and solutions.” Chemometrics and Intelligent Laboratory Systems 72(2): 209-217. Geladia, P., E. Bengtsson, K. Esbensen and H. Grahn (1992). ”Image- Analysis in Chemistry .1. Properties of Images, Greylevel Operations, the Multivariate Image.” Trac-Trends in Analytical Chemistry 11(1): 41-53. Geladib, P., H. Grahn, K. Esbensen and E. Bengtsson (1992). ”Image- Analysis in Chemistry .2. Multivariate Image-Analysis.” Trac- Trends in Analytical Chemistry 11(3): 121-130. Gellerstedt, G. and E.-L. Pettersson (1977). “Light-induced Oxidation of Lignin. Part 2. The Oxidative Degradation of Aromatic Rings. Svensk Papperstidning 80: 15-21. Ghosh, A. and Y. Ni (1998). ”Metal ion complexes and their relationship to pulp brightness.” Journal of Pulp and Paper Science 24(1): 26- 31. Gierer, J. and K.Jansbo (1993). ”Formation of Hydroxyl Radicals from Hydrogen Peroxide and their Effect on Bleaching of Mechanical Pulps.” Journal of Wood Chemistry and Technology 13(4):561. Givens, D.I., J.L. DeBoever and E.R. Deaville (1997). ”The principles, practices and some future applications of near infrared spectroscopy for predicting the nutritive value of foods for animals and humans.” Nutrition Research Reviews 10: 83-114. Gonzalez R.C., R.E Woods (2002). Digital Image Processing (2nd Edition). Prentice Hall, New Jersey. Grahn, H. F. and J. Saaf (1992). ”Multivariate Image Regression and Analysis - Useful Techniques for the Evaluation of Clinical Magnetic-Resonance Images.” Chemometrics and Intelligent Laboratory Systems 14(1-3): 391-396.

-64- References

Grgic, S., M. Grgic and B. Zovko-Cihlar (2001). ”Performance analysis of image compression using wavelets.” Ieee Transactions on Industrial Electronics 48(3): 682-695. Hagman, P.O.G. (1993) Automatic Quality Sorting of Picea-Abies Logs with a Gamma-Ray Log Scanner. Scandinavian Journal of Forest Research 8(4):583-590 Haralick, R.M., K.Shanmugam, I. Dinstein (1973). ”Textural features for image classification.” IEEE Trans. on System, man, and Cybernetics 3(6):610-621. Hauksson, J.B., G. Bergqvist, U. Bergsten, M. Sjöström and U. Edlund (2001). ”Prediction of basic wood properties for Norway spruce. Interpretation of Near Infrared Spectroscopy data using partial least squares regression.” Wood Science and Technology 35(6): 475-485. Henkereed, M.B. and S.N.C. Cheng (1993). ”Cloth Texture Classification Using the Wavelet Transform.” Journal of Imaging Science and Technology 37(6): 610-614. Joliffe, I.T. (1986). Principal Component Analysis. Springer. New York. Jordan, R., F. Feeney, N. Nesbitt and J.A. Evertsen (1998). ”Classification of wood species by neural network analysis of ultrasonic signals.” Ultrasonics 36(1-5): 219-222. Kelley, S.S., T.G. Rials, L.R. Groom and C.L. So (2004). ”Use of near infrared spectroscopy to predict the mechanical properties of six softwoods.” Holzforschung 58(3): 252-260. Kilroy, W.P. (1980). “Anaerobic decomposition of sodium dithionite in alkaline solution.” J. inorg. nucl. Chem. 42: 1071-1073. Kulcke, A., C. Gurschler, G. Spock, R. Leitner and M. Kraft (2003). ”On- line classification of synthetic polymers using near infrared spectral imaging.” Journal of near Infrared Spectroscopy 11(1): 71-81. Kvaal, K., J.P. Wold, U.G. Indahl, P. Baardseth and T. Naes (1998). ”Multivariate feature extraction from textural images of bread.” Chemometrics and Intelligent Laboratory Systems 42(1-2): 141- 158. Laine, A. and J. Fan (1993). ”Texture Classification by Wavelet Packet Signatures.” Ieee Transactions on Pattern Analysis and Machine Intelligence 15(11): 1186-1191. Lennholm, H. and T. Iversen (1995). ”Classification of Pulp Fibers from Different Wood Species by Multivariate Data-Analysis of C-13- Cp/Mas-Nmr-Spectra.” Holzforschung 49(5): 462-464.

-65- References

Lestander, T.A. and P. Geladi (2003). ”NIR spectroscopic measurement of moisture content in Scots pine seeds.” Analyst 128(4): 389-396. Lindgren, F. and P. Geladi (1992). ”Multivariate Spectrometric Image- Analysis - an Illustrative Study with 2 Constructed Examples of Metal-Ions in Solution.” Chemometrics and Intelligent Laboratory Systems 14(1-3): 397-412. Lister, M.W. and R.C. Garvie (1959). “Sodium dithionite, decomposition in aqueous solution and in the solid state.” Can. J. Chem. 37: 1567-1574. Liukko, K. (1997). Climate-adapted wet storage of saw timber and pulpwood : an alternative method of sprinkling and its effect on freshness of roundwood and environment. Doctoral thesis. Swedish University of Agricultural Sciences, SLU, Uppsala. Lorås, V. (1974). “Bleachability of Mechanical Pulp.” Tappi Journal 57(2):98. Lundstedt, T., E. Seifert, L. Abramo, B. Thelin, A. Nystrom, J. Pettersen and R. Bergman (1998). ”Experimental design and optimization.” Chemometrics and Intelligent Laboratory Systems 42(1-2): 3-40. Luo, J.B., D.W. Chen, K.J. Parker and T.S. Huang (1997). ”A scene adaptive and signal adaptive quantization for subband image and video compression using wavelets.” Ieee Transactions on Circuits and Systems for Video Technology 7(2): 343-357. Mallat, S. and W.L. Hwang (1992). ”Singularity Detection and Processing with Wavelets.” Ieee Transactions on Information Theory 38(2): 617-643. Martens, H., J.P. Nielsen and S.B. Engelsen (2003). ”Light scattering and light absorbance separated by extended multiplicative signal correction. Application to near-infrared transmission analysis of powder mixtures.” Analytical Chemistry 75(3): 394-404. Massart, D.L. and Y.V. Heyden (2004). ”From tables to visuals: Principal component analysis, part 1.” Lc Gc Europe 17(11): 586-+. Massart, D.L. and Y.V. Heyden (2005). ”From tables to visuals: Principal component analysis, part 2.” Lc Gc Europe 18(2): 84-+. McCaig, T.N. (2002). ”Extending the use of visible/near-infrared reflectance spectrophotometers to measure colour of food and agricultural products.” Food Research International 35(8): 731- 736. McClure, W.F. (2003). ”204 years of near infrared technology: 1800- 2003.” Journal of near Infrared Spectroscopy 11(6): 487-518. Michell, A.J. and L.R. Schimleck (1996). ”NIR spectroscopy of woods from Eucalyptus globulus.” Appita Journal 49(1): 23-26.

-66- References

Montgomery, D.C. (1997). Design and Analysis of Experiments, 4th Edition. John Wiley and Sons. New York. Myers, R.H. and D.C. Montgomery. (1995). Response Surface Methodology. John Wiley and Sons. New York. Niemz, P., F. Dutschmann and B. Stolken (1994). ”Using Nir- Spectroscopy for Evaluating Resin Content in Chip Mixtures.” Holz Als Roh-Und Werkstoff 52(1): 6-8. Norris, K.H. and J.D. Rowan (1962). “Automatic detection of blood in eggs.” Agricultural Engineering (St Joseph, Michigan) 43: 154-. Nystrom, J. and E. Dahlquist (2004). ”Methods for determination of moisture content in woodchips for power plants - a review.” Fuel 83(7-8): 773-779. Ohanian, P.P. and R.C. Dubes (1992). ”Performance Evaluation for 4 Classes of Textural Features.” Pattern Recognition 25(8): 819- 833. Oja, J., L. Wallbacks, S. Grundberg, E. Hagerdal and A. Gronlund (2003). ”Automatic grading of Scots pine (Pinus sylvestris L.) sawlogs using an industrial X-ray Log Scanner.” Computers and Electronics in Agriculture 41(1-3): 63-75. Osborne, B.G., T. Fearn and P.H. Hindle (1993) Practical NIR spectroscopy with applications in food and beverage analysis, 2nd edn. Longman Scientific & Technical. Harlow. Pearson, K. (1901) “On lines and planes of closest fit to systems of points in space.” Phil. Mag. 2: 559–72. Persson, E. (2001). Storage of Spruce Pulpwood – Effects on wood and mechanical pulp. Doctoral thesis. Swedish University of Agricultural Sciences, SLU, Uppsala. Persson, E. and J. Lindeberg (2001). ”Changes in metal concentrations during dry- and wet-storage of Norway spruce pulpwood.” Nordic Pulp & Paper Research Journal 16(4): 327-332. Persson, E. and T. Elowson (2001). ”Moisture content and discoloration of wood during dry and wet storage of Norway spruce (Picea abies (L.) Karst.) pulpwood.” Paperi Ja Puu-Paper and Timber 83(2): 132-137. Persson, E., J. Bergquist, T. Elowson, J. Jakara and B. Lonnberg (2002). ”Brightness, bleachability and colour reversion of groundwood made of wet- and dry-stored Norway spruce (Picea abies) pulpwood.” Paperi Ja Puu-Paper and Timber 84(6): 411-415. Poke, F.S., J.K. Wright and C.A. Raymond (2004). ”Predicting extractives and lignin contents in Eucalyptus globulus using near

-67- References

infrared reflectance analysis.” Journal of Wood Chemistry and Technology 24(1): 55-67. Rinker, R.G., S. Lynn, D.M. Mason and W.H. Corcoran (1965). “Kinetics and mechanism of the thermal decomposition of sodium dithionite in aqueous solution.” Industrial and Engineering Chemistry. Fundamentals 4(3): 282-288. Schimleck, L.R., A.J. Michell and P. Vinden (1996). ”Eucalypt wood classification by NIR spectroscopy and principal components analysis.” Appita Journal 49(5): 319-324. Sjöström M., S. Wold, W. Lindberg, J-A. Persson and H. Martens. (1983). “A multivariate calibration problem in analytical chemistry solved by partial least-squares models in latent variables.” Anal. Chim. Acta 150: 61-70. Sjöström, M., S. Wold and B. Söderström (1986). “PLS discriminant plots in Pattern Recognition in Practice II.” Elsevier Science. Publ. B. V. Holland. pp 461-470. So, C.L., B.K. Via, L.H. Groom, L.R. Schimleck, T.F. Shupe, S.S. Kelley and T.G. Rials (2004). ”Near infrared spectroscopy in the forest products industry.” Forest Products Journal 54(3): 6-16. Svensson, E., H. Lennholm and T. Iversen (1998). ”Pulp bleaching with dithionite: Brightening and darkening reactions.” Journal of Pulp and Paper Science 24(8): 254-259. Svensson, V.T., H.H. Nielsen and R. Bro (2004). ”Determination of the protein content in brine from salted herring using near-infrared spectroscopy.” Lebensmittel-Wissenschaft Und-Technologie-Food Science and Technology 37(7): 803-809. Thygesena, L.G. and S.O. Lundqvist (2000). ”NIR measurement of moisture content in wood under unstable temperature conditions. Part 1. Thermal effects in near infrared spectra of wood.” Journal of near Infrared Spectroscopy 8(3): 183-189. Thygesenb, L.G. and S.O. Lundqvist (2000). ”NIR measurement of moisture content in wood under unstable temperature conditions. Part 2. Handling temperature fluctuations.” Journal of near Infrared Spectroscopy 8(3): 191-199. Tredway, C.M. (1979). “Factors affecting the bleaching of groundwood with sodium hydrosulfite.” March Issue 73-75. Trygg, J. and S. Wold (2002). ”Orthogonal projections to latent structures (O-PLS).” Journal of Chemometrics 16(3): 119-128. Trygg, J., N. Kettaneh-Wold and L. Wallbacks (2001). ”2D wavelet analysis and compression of on-line industrial process data.” Journal of Chemometrics 15(4): 299-319.

-68- References

Tsuchikawa, S. and K. Yamato (2003). ”Discriminant analysis of wood- based materials with weathering damage by near infrared spectroscopy.” Journal of near Infrared Spectroscopy 11(5): 391- 399. Tsuchikawa, S., M. Torii and S. Tsutsumi (1996). ”Application of near infrared spectrophotometry to wood .4. Calibration equations for moisture content.” Mokuzai Gakkaishi 42(8): 743-754. Unser, M. (1995). ”Texture Classification and Segmentation Using Wavelet Frames.” Ieee Transactions on Image Processing 4(11): 1549-1560. Van de Wouwer, G., P. Scheunders, S. Livens and D. Van Dyck (1999). ”Wavelet correlation signatures for color texture characterization.” Pattern Recognition 32(3): 443-451. Vanespen, P., G. Janssens, W. Vanhoolst and P. Geladi (1992). ”Imaging and Image-Processing in Analytical-Chemistry.” Analusis 20(3): 81-90. Wallbacks, L., U. Edlund, B. Norden and I. Berglund (1991). ”Multivariate Characterization of Pulp Using Solid-State 13cnmr, Ftir, and Nir.” Tappi Journal 74(10): 201-206. Wang, D.J., X.Y. Zhou, T M. Jin, X.N. Hu, J.E. Zhong and Q.T. Wu (2004). ”Application of near-infrared spectroscopy to agriculture and food analysis.” Spectroscopy and Spectral Analysis 24(4): 447-450. Wilhelmsen, G. (1968). “Water storage of unbarked mechanical pulp wood of Norway Spruce.” Norsk Skogsindustri 6:207-211. Wold, S. (1978). “Cross validatory estimation of the number of components in factor and principal components models.” Technometrics 20:397-405. Wold, S., H. Antti, F. Lindgren and J. Ohman (1998). ”Orthogonal signal correction of near-infrared spectra.” Chemometrics and Intelligent Laboratory Systems 44(1-2): 175-185. Wold, S., K. Esbensen and P. Geladi (1987). ”Principal Component Analysis.” Chemometrics and Intelligent Laboratory Systems 2(1- 3): 37-52. Yeh, T.F., H.M. Chang and J.F. Kadla (2004). ”Rapid prediction of solid wood lignin content using transmittance near-infrared spectroscopy.” Journal of Agricultural and Food Chemistry 52(6): 1435-1439. Yu, H. L. and J. F. MacGregor (2003). ”Multivariate image analysis and regression for prediction of coating content and distribution in the

-69- References

production of snack foods.” Chemometrics and Intelligent Laboratory Systems 67(2): 125-144.

-70- Acknowledgements

12. Acknowledgements

Först och främst vill jag tack min huvudhandledare Ulf Edlund för den kloka vägledning jag fått, den noggrannhet du haft när du läst mina artiklar och den förståelse jag fått när jag behövt det. Jag vill även tacka min biträdande handledare Michael Sjöström för dina intelligenta kemometriska synpunkter, ibland när man allra minst anade det.

Vidare vill jag tacka Roland Agnemo, Domsjö Fabriker, för det engagemang du visade när jag utförde mina blekningsförsök på dåvarande Modo. Du ska också ha tack för att jag fått fiska ett flertal gånger i de fina representationsfiskevattnen!

Ett stort tack till min kära Josefina för att du stått ut med mig i ur och skur. Jag vill också tacka min mamma Berit, min pappa Bo och mina syskon Esbjörn och Olle för att ni finns där.

Sen vill jag tacka min gamle följeslagare och barndomsvän sen lågstadiet, Johan ”Pinocchio” Trygg. Tack för dina synpunkter och alla goda råd jag fått angående min forskning. Tack för att du kallar mig ”pucko”.

Tack till Erik Persson, SLU Uppsala, för dina idéer, ditt engagemang och för att du vidgat mina vyer när det gäller forskningen.

Tack till Kristina Elg Christoffersson, min rumskompis under många år. Dina tankar och frågeställningar har fått mig att tänka till och analysera saker som jag annars inte hade gjort.

Tack till Carina och Ann-Helèn för allt hjälp jag fått och för att ni sett till att allt flutit på.

Tack till Bert Larsson för din vänskap och våra diskussioner angående olika mest tekniska saker.

-71- Acknowledgements

Tack till Thomas Jonsson, Umetrics AB, för alla dom gånger Simca har jävlats och du kommit som räddaren i nöden.

Tack till mina gamla kompisar Per-Erik, Carl-Mikael, Stefan och Conny. Jag hoppas att vi fortfarande kan hålla kontakten och hitta på saker.

Tack till Tobias Stenström och Dan Johnels för er vänskap och ert engagemang angående foto. Roligt att ha någon att diskutera hobbies med på jobbet.

Jag vill också tacka fiskarna Ulf, Frippe, David, Mickael M, Mikael S, Hans och Tomas för alla trevliga fisketurer och allt fiskesnack.

Tack till Jon Gabrielsson för alla intressanta pryldiskussioner. Tack till Freddan för du hållt uppe mitt musik och dataspelsintresse. Tack till Suss för våra diskussioner, hoppas vi ses mer i Innansjön! Tack till Anders B och Lissan. Ett stort tack till övriga kemometriker nu och då; Anton, Ing- Marie, Henrik, Anna, Max, Malin, Pär och Per. Jag hoppas jag inte glömde nån nu.

Ett stort tack till övriga personer på Organisk Kemi som jag inte nämnt med namn!

Tack till Agne, Eva, Hjalmar, Julia, Johanna, Karin, Daniel, Helena och Markus för att ni gör det lättare att vara ute i obygden.

Tack Bim, Isa och Claes!

-72-

Classification of the Softwood Species Spruce and Pine using Near-infrared Reflectance Measurements on Bark combined with Multivariate Data Analysis

David Nilssona,* and Erik Perssonb aDepartment of Chemistry, Organic Chemistry, Umeå University, SE-901 87 Umeå bDepartment of Forest Products and Markets, P.O. Box 7060, SE-750-07 Uppsala

Abstract

Near-infrared reflectance (NIR) measurements were performed on the bark for a number of 25 discs of Norway Spruce and 28 discs of Scots Pine. Principal Component Analysis (PCA) was applied to the spectral data in order to visually classify the pine and spruce objects. Partial

Least Squares (PLS) to latent structures were then used to find regression models between the collected spectra and the roundwood species classification of the measured discs. Separate test sets used for validation of the PLS models were selected from the spectral data and predicted.

It was found that PCA to some degree could be used to estimate the roundwood species. PLS gave more accurate results and provided a simpler solution for the roundwood classification.

All predicted objects were classified correctly using the PLS modelling. The results showed that measuring bark with NIR spectroscopy might be a viable method for automatic classification of roundwood.

*Author to whom correspondence should be addressed Email address: [email protected] Telephone: +46 90 786 5999 Fax: +46 90 13 88 85 Introduction

In general, different species of wood are used for different end products with different pricing and handling in the sawmill process. The roundwood delivered to the saw mills are still to this date delivered as mixed species from the forest. As a result, the different species must be separated at some stage in the production chain, i.e. from forest to sawn timber. This is also true, to some extent, for the quality grading. While different end products have different requirements of grade, the drying operations must be adapted to the end products. The sorting or separation of different species and quality grades takes place at different stages in the chain, which depend on the number of commercially available species, the structure of forests or industries, traditions and the available technology.

In North America the lumber is often delivered as mixed softwood species to the sawmills and the species or groups of species are separated after the sawing, prior to the drying operations.

In Sweden the two most important species used for timber production are Norway Spruce

(Picea abies (L.) Karst.) and Scots Pine (Pinus sylvestris). According to the Swedish sawmill inventory for year 2000, these species represent more than 98% of the production. The remaining 2% mainly consists of different hardwood species. The sawmills often use both

Scots Pine and Norway Spruce as raw material and the species are separated at the scaling operation prior to sawing. The classification of species and the quality grading is done by manual visual assessment of the logs. The situation is similar in other Scandinavian countries.

As a part of the general strive to increase the productivity and cut costs in the sawmill industry, efforts are made to develop automatic systems for log scaling. Methods for automatic measurement of length and diameter have been in use at sawmills for a long time. The sorting and grading of logs for payment to the supplier is usually made manually, but the use of different automatic methods is increasing. Techniques like gamma-ray (Hagman 1993),

X-ray (Oja et al. 2003; Grundberg and Gronlund 1997) and optical 3D (Gjerdum et al. 2001;

Oja et al. 2004) log scanning can facilitate the grading of logs. There are however very few automatic methods for species classification of logs described in the literature or patent databases. Existing methods for wood classification are instead based on sawn products and the majority of them are using a spectroscopic approach. Wood identification has been done by means of Fourier-Transform NIR (Brunner et al. 1996a; Brunner et al. 1996b), with Ion

Mobility Spectrometry (Lawrence 1989) and with Raman Spectrometry (Lewis et al. 1994;

Yang et al. 1999). There is also a method using ultrasonic signals (Jordan et al. 1998).

The present pilot study suggests a method for automatically classify wood species with near- infrared reflectance (NIR) bark measurements in combination with multivariate data analysis.

NIR spectroscopy can be used in industrial applications for quality control and for classification of raw material in the pharmaceutical- (Zhou et al. 1998; Columbano et al.

2002), food- (Wahlby and Skjoldebrand 2001) and pulp & paper (Wright et al. 1990;

Hauksson et al. 2001) industries. NIR spectroscopy gives information about both the chemical composition (Osbourne et al. 1993) as well as the physical properties (Osbourne et al. 1993) of a sample. The different species of wood differs both in chemical composition (Hon and

Shiraishi 1991), wood anatomy and colour. For unbarked logs, the different properties of the bark could be utilised for classification. Since the wavelengths of any NIR spectrum are highly correlated and consists of numerous underlying peaks (Osbourne et al. 1993), multivariate data analysis methods such as Principal Component Analysis PCA (Wold et al.

1987) and Partial Least Squares (PLS) regression to latent structures (Geladi et al. 1986;

Hoskuldsson 1988) must be applied to the spectral data. Prior to calculating multivariate models, the spectra are often processed with some kind of filtering in order to reduce unwanted variation which facilitates the modelling. Multiplicative Scatter Correction (MSC)

(Geladi et al. 1985) and Orthogonal Signal Correction (OSC) (Wold et al. 1998) are two pre- treatment methods viable for filtering NIR spectra.

Materials and methods

The used wood samples were discs of Scots Pine (Pinus sylvestris) and Norway Spruce (Picea abies (L.) Karst.) collected in a pile of logs from a thinning made approximately one month before the collection of samples. There were totally 53 discs, 25 of Norway spruce and 28 of

Scots pine.

The measurements were performed using a NIR Systems 6500 Spectrophotometer equipped with a fibre-optic probe. Measurements were done both on the bark and on the sapwood.

During the measurements the probe was held against the surface of the bark or the radial wood surface and the reflectance was measured. The sapwood measurements were accomplished by putting the probe against the sawn surface of the disc. The radial surface had the original properties created by chainsaw during the collection of samples.

Reflectance spectra of the 53 samples (48 sapwood samples) were collected in the wavelength range 400-2500 nm with a spectral resolution of 2 nm. The scanned spectra were assembled into two data matrices (53 objects and 1050 variables), one for the bark measurements and one for the sapwood measurements.

Multivariate PCA and PLS models were calculated from raw, MSC-treated and OSC-treated spectra. MSC pre-treatment implies that the spectra undergo a regression towards the average spectrum and that each spectrum is corrected with respect to both baseline shift and slope.

The use of OSC is only feasible when a response variable is available. The OSC treatment removes variation in the X matrix (NIR spectra) that is not correlated to the chosen response or y-variable. In this study the y-variable was the species classification of the roundwood, i.e. spruce or pine.

PCA was used to describe the multidimensional spectral variable space with a few significant principal components that covered most of the variation. More generally, principal components are comprised of scores and loadings, which, if they are plotted, can give valuable information about the samples and variables in the data set. A score plot consists of the score values for one up to three components and shows how the similarities or differences between samples or objects. Objects with similar properties are lying closely to each other and can often form a cluster. A loading plot is comprised of the loading values for up to three principal components and reveals what variables or wavelengths that are important for the separation of objects in the score plot.

PLS was used to find correlation between the spectral data and the roundwood species classification variable. In the general case, PLS finds latent structures, or components, in the

X matrix and correlates these with components in the Y matrix. Often, the Y matrix consists of only one variable, y. The data are usually divided into a calibration set that is used for model building, and a test set that is used for validation. In order to obtain a good prediction of the test set, an optimal number of PLS components has to be calculated. Too few components will result in poor prediction due to the fact that the under-fitted model has not included variation important for the correlation to the Y matrix. Too many components will cause the model to become over-fitted and the superfluous components will mostly be comprised of noise that will lower the prediction ability of the model. In this study, the cross- validated Q2Y-value (Wold 1978) was used to find the number of significant components.

Results and discussion

The NIR spectra collected from bark of roundwood contain considerable amounts of variation that is due to various scatter effects. These can arise from the properties of the measured log surface such as irregularities in the bark or by a varying distance between the measured log and the spectroscopic near-infrared probe. A plot (fig 1) of the fifty-three recorded spectra measured in this study shows that a baseline shift as well as a variation in slope can be found within the spectral data. The corresponding multiplicative scatter corrected spectra can be seen in fig 2, which shows that the spectral variation has been reduced significantly. The OSC treatment represents an even heavier reduction of the spectral variation and can be seen in fig

Classification of roundwood using Principal Component Analysis

Principal Component Analysis was applied to the multiplicative scatter corrected spectra of the measured roundwood. The wavelengths between 400 to 600 nm and 2250 to 2500 were removed prior to modelling in order to exclude random noise that was found in these regions.

The five-component model derived from the spectra explained 99% of the variation in data.

The score plot for the first three components (fig 4) shows that the objects are grouped into three different classes. The group in the middle corresponds to the spruce samples while the two other groups are comprised of the pine samples. The reason for the separation of the pine objects into two groups was that the measured logs had bark that differed with respect to the surface smoothness. The bark of a living pine tree is often very rough and dark-coloured at the base while the top of the tree has a brighter and smoother bark. The score plot for the second, third and forth components (fig 5) divides the objects into two classes, pine and spruce. It is noticeable that in both score plots there are objects that lie in the wrong class. The objects in the spruce class are grouped tightly together, so the faulty classification is mainly caused by the large spread of the pine samples. It is evident that much of the variation in the principal components is due to the structure of the tree bark where the bark of spruce is more homogenous. A PCA model was also calculated for the raw centred spectra but the score plot for the first three components (not shown) had a worse separation between the objects compared to the model using MSC treated spectra.

Based on the score plot for the first three components, two test sets of 15 objects each that spanned the variation in data with respect to both structure and chemical properties were selected. To ensure that representative test sets were used, the objects that were dubiously classified in the score plots were not discriminated and several of these objects were included in the test sets. The test sets were then used for external validation of the models created in this study. In order to further examine the classification characteristics of the PCA modelling,

Cooman (Coomans et al. 1984) classification plots were created for the two test sets. In order to exclude variation that originated from light scattering, the models were calculated from

MSC-treated spectra. The concept behind making these plots was that one PCA model was made for each of the three groupings seen in the first score plot. It is noticeable that two models had to be made for the pine objects since the variation within these objects was too large. After that, the distance between the objects in the test sets and the three models could be plotted. A short distance to one of the models implies that a certain object belongs to that particular class and thus the object can be classified as either spruce or pine (rough or smooth bark). Of course, separate models had to be made for the two test sets, since the corresponding calibration sets were consisted of different samples. The results from the Cooman classification of the first test set (fig 6 a,b,c) were rather ambiguous since several of the pine samples with smooth bark were predicted erroneously. The two classes comprised of spruce samples and pine samples with rough bark were however classified almost correctly. The

Cooman classification for the second test (not shown) set showed similar results compared to the first set.

Classification of roundwood using Partial Least Squares regression to latent structures

Partial Least Squares Discriminant Analysis (PLS-DA) (Sjostrom et. Al) models using the

NIR spectra as X matrix and the roundwood species classification as y-variable were developed. The y-variable consisted of a single column that was set either at “0” or “1”, denoting pine or spruce class belonging. Prior to computing the regression models, the NIR spectra were filtered using Orthogonal-PLS (Trygg and Wold 2002) (O-PLS), a variant of the

Orthogonal Signal Correction algorithm. The O-PLS treatment implied that the NIR spectra were filtered from variation in the data that was not correlated to the roundwood species classification. During the modelling prior to using O-PLS, it was found that a fairly large number of latent variables was needed to describe the correlation between X and y. The O-

PLS managed to exclude the orthogonal variation within these components and PLS could then be used to fit a final one-component model between the NIR spectra and the species classification variable for each of the calibration sets. In any case, the same number of components had to be calculated whether O-PLS was used or not, but the O-PLS utilisation provided a greatly simplified solution for the final prediction models. The prediction of the test sets is shown in fig 7 where it can be seen that all excluded objects are predicted correctly.

A predicted value below 0.5 denotes pine class belonging, while a predicted value above 0.5 denotes spruce class belonging. The correlation between X and y is represented by the PLS loading weights (fig 8), which show that roundwood species classification correlates to whole used spectral region, but especially to the wavelengths around 800 nm which corresponds to the upper part of the visual region and lower part of the near-infrared region.

Moreover, PLS models between the NIR spectra and the roundwood species classification were calculated using reduced data matrices. Models that were comprised of the wavelength region between 600 to 1100 nm and models that only used the wavelengths between 1100 to

2250 nm were created. The modelling showed that is was possible to use either of the two wavelength regions and yet obtain very similar prediction results to when using the whole region. All predicted samples were classified correctly in the predicted vs observed plots (not shown). The PLS weights for the models using the wavelength region between 600 to 1100 nm indicated that the species classification, as showed earlier, is strongly correlated to wavelengths around 800 nm. The correlation to the high near-infrared region (fig 9) includes almost all wavelengths but especially those that are close to 2000 nm. The reader is referred to table 1 for a summary of the most important models built in this study.

Classification of roundwood from NIR spectra measured on sapwood surface

Multivariate models were also computed from spectra measured on the sawn wood surface of each disc. This was done in order to examine how well this modelling compared to the models created from spectra of bark. An initial PLS model using the whole spectral region between

400 and 2500 showed that the roundwood classification is much correlated to lower part of the visual region (fig 10). A PCA model based on MSC treated spectra using the wavelengths between 400 to 2250 nm was calculated and the score plot for the first three components (fig

11) reveals that it is hard to distinguish the spruce samples from the pine samples. One possible outlier was also detected and removed before further modelling. A test set consisting of 16 samples was excluded from the calculation of a final PLS model that was comprised of the same wavelengths used in the PCA model. The prediction of the test set (fig 12) showed that roundwood classification was estimated fairly accurately with two objects being wrongly classified.

Conclusions

This study has shown that PCA can be applied to spectra collected from bark in order to some degree classify spruce and pine roundwood. Much of the variation in the spectra is due to light scatter effects, which are caused by irregularities of the measured roundwood. The scattering is partly correlated to the measured roundwood species and much of the irrelevant variation can be corrected using data pre-treatment methods such as Multiplicative Scatter Correction or Orthogonal Signal Correction. It was not possible, however, to predict the roundwood species classification using Cooman classification since the pine objects with smooth bark did not form a well defined class. In any case, it is evident that the use of PLS regression will allow for a much safer prediction of the roundwood species. The PLS modelling will find components that are comprised of variation that arises from structural as well as chemical differences in the measured material. This is not the case with the PCA modelling, where the significant components mainly consists of variation due to various scattering effects. Another advantage of using PLS instead of PCA is that the classification is much easier. The utilisation of OPLS will provide parsimonious PLS models that can be used for reliable predictions of the roundwood species.

Compared to the prediction from bark, the wood measurements failed to classify the roundwood samples when using PCA. The classification using PLS was however much better and proved to be almost as good as the bark classification. Though it is evident that there is a larger separation between spruce and pine bark than between spruce and pine wood that is due to both chemical and structural differences, the wood classification might prove to be useful in those cases the bark is not present on the measured log.

The final conclusions that can be made is that the bark classification worked well under laboratory conditions. There are however a number of disturbing factors such as temperature, background light, distance between log and measuring probe and finally the limited time for data collection that need to be taken into account in a possible industrial application. In order to handle these factors, a robust calibration model that consists of a variety of samples that cover all possible sample variation needs to be assembled. This may sound cumbersome, but the advantage from using NIR spectroscopy is that the model creation will be a swift process due to the speed of the near-infrared measurements combined the simplicity of the reference variable. The outcome of this study suggests that the present method is not only applicable to the spruce and pine species, but to other species as well, both coniferous and decidous.

Compared to other techniques, such as X-ray, NIR spectroscopy should prove itself to be a feasible method due its low-cost, speed and predictive ability when combined with multivariate data analysis.

Acknowledgements

Grants from The Swedish Council for Forestry are gratefully acknowledged.

References to the literature

Brunner C, Maristany A, Butler D (1996) Wood species identification using spectral reflectance. Forest Products Journal 46(2):82-85 Brunner M, Eugster R, Trenka E, Bergamin-Strotz L (1996) FT-NIR Spectroscopy and wood identification. Holzforschung 50:130-134 Columbano A, Buckton G, Wikely P (2002) A study of the crystallisation of amorphous salbutamol sulphate using water vapour sorption and near infrared spectroscopy. International Journal of Pharmaceutics 237:171-178 Coomans D, Broeckaert I, Derde MP, Tassin A, Massart DL, Wold S (1984) Use of a microcomputer for the definition of multivariate confidence regions in medical diagnosis based on clinical laboratory profiles. Computers and Biomedical Research 17(1):1-14 Geladi P, Kowalski BR (1986) Partial Least Squares Regression: A Tutorial. Analytica Chimica Acta 185:1-17 Geladi P, MacDougall D, Martens H (1985) Linearization and Scatter-Correction for Near- Infrared Reflectance Spectra of Meat. Applied Spectroscopy 39(3):491-500 Gjerdum P, Warensoj M, Nylinder M (2001) Classification of crook types for unbarked Norway spruce sawlogs by means of a 3D log scanner. Holz als Roh-und Werkstoff 59(5):374-379 Grundberg S, Gronlund A (1997) Simulated grading of logs with an X-ray log scanner – Grading accuracy compared with manual grading. Scandinavian Journal of Forest Research 12(1):70-76 Hagman POG (1993) Automatic sorting of Picea abies logs with a Gamma-ray log scanner. Scandinavian Journal of Forest Research 8(4):583-590 Hauksson JB, Bergqvist G, Bergsten U, Sjostrom M, Edlund U (2001) Prediction of basic wood properties for Norway spruce. Interpretation of Near Infrared Spectroscopy data using partial least squares regression. Wood Science and Technology 35:475-485 Hon DNS, Shiraishi N (1991) Wood and cellulosic chemistry. Marcel Dekker, New York Hoskuldsson A (1988) PLS regression Methods. Journal of Chemometrics 2:211-228 Jordan R, Feeney F, Nesbitt N, Evertsen JA (1998) Classification of wood species by neural network analysis of ultrasonic signals. Ultrasonics 36:219-222 Lawrence AH (1989) Rapid Characterization of Wood Species by Ion Mobility Spectrometry. Journal of Pulp and Paper Science 15(5):196-199 Lewis I, Daniel N, Chaffin N, Griffiths P (1994) Raman Spectrometry and neural networks for the classification of wood types. Spectrochimica Acta Part A: Molecular Spectroscopy 50(11):1943-1958 Oja J, Wallbacks L, Grundberg S, Hagerdal E, Gronlund A (2003) Automatic grading of Scots pine (Pinus sylvestris L.) sawlogs using an industrial X-ray log scanner. Computers and Electronics in Agriculture 41(1-3):63-75 Oja, J, Grundberg S, Fredriksson J, Berg P (2004) Automatic grading of sawlogs: A comparison between X-ray scanning, optical three dimensional scanning and combinations of both methods. 19(1):89-95 Osbourne BG, Fearn T, Hindle PH (1993) Practical NIR spectroscopy with applications in food and beverage analysis, 2nd edn. Longman Scientific & Technical, Harlow Sjostrom M, Wold S, Soderstrom B (1986) PLS discriminant plots. In: Pattern recognition in practice II. ES Gelsema and LN Kanal, Eds, Elsevier, Amsterdam, 486. Trygg J, Wold S (2002) Orthogonal projections to latent structures. Journal of Chemometrics 16:119-128 Wahlby U, Skjoldebrand C (2001) NIR-measurements of moisture changes in foods. Journal of Foof Engineering 47:303-312 Wold S (1978) Cross validatory estimation of the number of components in factor and principal components models. Technometrics 20:397-405 Wold S, Antti H, Lindgren F, Ohman J (1998) Orthogonal Signal Correction of Near-Infrared Spectra. Chemometrics and Intelligent Laboratory Systems 44:175-185 Wold S, Esbensen K, Geladi P (1987) Principal Component Analysis. Chemometrics and Intelligent Laboratory Systems 2:37-52 Wright JA, Birkett MD, Gambino, MJT (1990) Prediction of pulp yield and cellulose content from wood samples using near infrared reflectance spectroscopy. Tappi Journal Yang H, Lewis I, Griffiths P (1999) Raman spectrometry and neural networks for the classification of wood types. 2. Kohonen self-organizing maps. Spectrochimica Acta Part A 55:2783-2791 Zhou X, Hines P, Borer MW (1998) Moisture determination in hygroscopic drug substances by near infrared spectroscopy. Journal of Pharmaceutical and Biomedical Analysis 17:219- 225

Figure and table captions

Table 1. Summary of the multivariate models calculated in this study. aModel Type. bSamples used for bulding the model. Bark=bark samples, Sapwood=sapwood samples, PR=Pine with rough bark, PS=pine with smooth bark, S=spruce, CS1=calibration set 1, CS2=calibration set 2, CSS=calibration set spruce. cWavelength region used (nm). dPre-treatment method used, Ctr=centring, MSC=Multiplicative Scatter Correction, OPLS=Orthogonal PLS. eNumber of components for calibration model. fNumber of components calculated for OPLS. gVariation in X removed by OPLS filtering. hVariation in X explained by model. iCross-validated value for internal model validation. jPer cent of samples correctly classified.

Figure 1. Raw near-infrared spectra of bark. The spectra exhibit a baseline shift as well as differences in slope.

Figure 2. MSC treated near-infrared spectra of bark. The MSC filtering has reduced the variation within the spectra significantly.

Figure 3. OSC treated spectra of bark. The OSC filtering has removed around 95% of the spectral variation.

Figure 4. Score plot for the first three components of the PCA model made for the near- infrared and MSC treated spectra of bark. Spruce samples are marked with spheres and pine samples with cubes. Pine samples with rough bark are found to the left and pine samples with smooth bark to the right.

Figure 5. Score plot for component two, three and four of the PCA model made for the near- infrared spectra of bark. Spruce samples are marked with spheres and pine samples with cubes.

Figure 6. Cooman classification plots for the three PCA models based on spectra of bark. a) Normalised distances to the models made for spruce and pine with rough bark. Three test set objects are correctly classified as pine with rough bark and seven objects are correctly classified as spruce. One spruce object, no. 8, is wrongly classified. The pine objects with smooth bark are classified to neither of the groups. The classification used a significance level of 95%, which is represented by the vertical and horizontal line in the plot. b) Normalised distances to models made for spruce and pine with smooth bark. All spruce objects with the exception of no. 8 are classified correctly. One pine object with smooth bark, no. 32, is almost classified correctly. The rest of the pine objects are classified to neither of the models using a significance level of 95%. c) Normalised distances to models made for pine with smooth and rough bark. The three pine objects with rough bark are classified correctly. The rest of the objects are belonging to neither of the models according to a significance level of 95%. Figure 7. Observed vs. predicted plot for the excluded objects in the two test sets. The objects were predicted using the PLS calibration models calculated from the spectra of bark. The wavelength region between 600 and 2250 nm were used in the modelling. All predicted objects were classified correctly.

Figure 8. Loading weights for the one-component PLS model created from spectra of bark in the wavelength region. The loading weights represent the wavelengths in X that correlate to y. A strong correlation was found to the wavelengths around 800 nm.

Figure 9. Loading weights for the one-component PLS model calculated from spectra of bark in the wavelength region of 1100 to 2250 nm. The correlation to y is particularly described by the wavelengths around 2000 nm.

Figure 10. Loading weights for the PLS model calculated from spectra collected from sapwood. The wavelengths in the visual region are strongly correlated to the roundwood species classification.

Figure 11. Score plot for the first three components of the PCA model computed from multiplicative scatter corrected spectra measured on sapwood. Apart from one outlier, the objects are not separated into different classes.

Figure 12. The prediction of roundwood for the excluded test set during the modelling of spectra collected from sapwood measurements. Two spruce objects are classified incorrectly.

Figure 1.

1.2

0.8

0.6 log 1/R

0.4

0.2

0 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm)

Figure 2.

0.9

0.8

0.7

0.6 log 1/R

0.5

0.4

0.3

0.2 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm)

Figure 3.

0.9

0.8

0.7

0.6 log 1/R

0.5

0.4

0.3

0.2 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm)

Figure 4.

Figure 5.

Figure 6. a) 14

12 42 32 10

8 30 38

Spruce 6 2 19 4 17 2 8 20 41 3325 37 4547 0 0123456789101112131415 Pine Rough b)

12 42

32 10

8 30 38 2 6 Spruce 19 4 17

2 8 4120 2547 3345 37 0 0123456789101112 Pine Smooth c)

8 37

45 19 30 42 Pine Smooth 6 33

4 17 382 8 20 41 2 47 25 32

0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Pine Rough Figure 7.

0.8

0.6

Observed 0.4

0.2

-0.2 -1.5 -1 -0.5 0 0.5 1 1.5 Predicted

Figure 8.

0.07

0.06

0.05

0.04

PLS Weights w1* Weights PLS 0.03

0.02

0.01 600 800 1000 1200 1400 1600 1800 2000 2200 Wavelength (nm)

Figure 9.

0.055

0.05

0.045

0.04

PLS Weights w1* Weights PLS 0.035

0.03

0.025 1200 1400 1600 1800 2000 2200 Wavelength (nm)

Figure 10.

0.1

0.08

0.06

0.04

0.02

-0.02 PLS Weights w1* Weights PLS -0.04

-0.06

-0.08

-0.1 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm)

Figure 11.

Figure 12.

0.8

0.6

Observed 0.4

0.2

-0.2 -0.5 0 0.5 1 Predicted

Table 1.

a b c d e f g 2 h 2 i j M O W P A AOPLS XVortho R X Q y C PCA Bark 600-2250 Ctr, MSC 5 - - 0.99 - - PCA Bark, PR, CS1 600-2250 Ctr, MSC 3 - - 0.98 - 67 % PCA Bark, PS, CS1 600-2250 Ctr, MSC 5 - - 1.00 - 67 % PCA Bark, S, CS1 600-2250 Ctr, MSC 6 - - 1.00 - 67 % PLS Bark, CS1 600-2250 Ctr, OPLS 1 14 93.31 % 1.00 0.758 100 % PLS Bark, CS2 600-2250 Ctr, OPLS 1 14 99.02 % 1.00 0.737 100 % PLS Bark, CS1 600-1100 Ctr, OPLS 1 11 83.22 % 1.00 0.727 100 % PLS Bark, CS2 600-1100 Ctr, OPLS 1 11 94.55 % 1.00 0.673 100 % PLS Bark, CS1 1100-2250 Ctr, OPLS 1 12 95.79 % 1.00 0.772 100 % PLS Bark, CS2 1100-2250 Ctr, OPLS 1 12 99.95 % 1.00 0.832 100 % PCA Sapwood, CSS 400-2250 Ctr, MSC 6 - - 0.99 - - PLS Sapwood, CSS 400-2250 Ctr, OPLS 1 6 98.69 % 0.99 0.541 87.5 %

Holzforschung, Vol. xx, pp. xxx-xxx, 2005 • Copyright by Walter de Gruyter • Berlin • New York. DOI 10.1515/HF.2005.XXX

Pine and spruce roundwood species classification using multivariate image analysis on bark

David Nilsson* and Ulf Edlund such as X-ray scanning (Oja et al. 2003), gamma-ray scanning (Hagman 1993) and 3D log scanning (Gjerdrum Department of Chemistry, Organic Chemistry, Umea˚ et al. 2001). It is unclear at present if any of these meth- University, SE-901 87 Umea˚ , Sweden ods can be used for automatic roundwood species clas- *Corresponding author. sification. Sawn wood species, on the other hand, have E-mail: [email protected] been classified by a number of methods, including FT- NIR spectroscopy (Brunner et al. 1996), Raman spectroscopy (Lewis et al. 1994), ion mobility spectrometry Abstract (Lawrence 1989) and ultrasonic signals (Jordan et al. 1998). An automated method has also been proposed for Wood discs from 67 pine and 79 spruce logs were col- detecting defects in sawn wood using image segmen- lected from a forest clearing. Three different 24-bit red- tation (Funck et al. 2003). Again, it is not known if the green-blue (RGB) images were acquired from the radial methods available for sawn wood can be used for auto- surface of each disc. The first image contained bark, the matic roundwood species classification. second image was a mixture of bark and wood surface, and the third image consisted only of wood surface. The Multivariate image analysis image texture was compressed into vectors of Fourier- transformed wavelet coefficients. These were assembled A multivariate digital image is a three-way data table in in matrices and analysed by principal component ana- which the two first dimensions represent the phys- lysis (PCA) and partial least-squares projections to latent ical size and the third dimension, the number of colour structures (PLS). Classification using Fourier-transformed channels or wavelengths measured. A grey-scale image wavelet coefficients showed that the wood species could reduces down to a 2D matrix, as the third dimension con- be predicted with 90% accuracy. A thorough examination sists of a single channel. A red-green-blue (RGB) image of this classification showed that the predicting power of (Gonzalez and Woods 2002) can consequently be viewed these models was mostly due to wavelet coefficients that as a three-dimensional array (N,K,Is3), where N is the represented the mean value of each colour channel. The pixel rows, K the pixel columns and I the number of col- prediction accuracy that could be obtained from coeffi- our channels, which in this case is three. Traditional cients representing image texture was generally low. The spectral analysis uses histograms to bin the image pixels use of grey-level co-occurrence matrices prior to the according to their intensities in each colour channel wavelet transformation showed, however, that it is pos- (Gonzalez and Woods 2002). While histogram analysis is sible to classify the wood species of pine and spruce a more powerful method than just looking at the mean with an accuracy approaching 100%. values for each channel, spatial information is lost. Therefore, analysis of texture can be used as a comple- Keywords: bark; co-occurrence matrix; multivariate ment to colour evaluation. Combinatorial image analysis image analysis; partial least squares projection to latent of colour and texture at the same time is, however, not structures (PLS); pine; spruce; wavelet; wood species. necessary (Ma¨ enpa¨a¨ and Pietikainen 2004). A compre- hensive review (Bharati et al. 2004) of different methods for image texture analysis has been published. Further- Introduction more, analysis of image textures has been used to analyse and classify various forms of food, including textural Norway spruce (Picea abies) and Scots pine (Pinus syl- analysis of bread (Kvaal et al. 1998), evaluation of food vestris) are the two most common wood species in Swe- colour (Antonelli et al. 2004) and classification of bovine den and Scandinavia and account for nearly all wood that meat (Basset et al. 2000). The former two studies both is transported to sawmills. The majority of the Swedish extracted image features assembled as matrices, which sawmills, i.e., 75% according to the Swedish sawmill were analysed by means of multivariate data analysis. inventory, process both Norway spruce and Scots pine. Unlike traditional multivariate image analysis (MIA) Even though there is rough division of the species during (Esbensen and Geladi 1989; Geladi et al. 1989), which felling and transportation, logs can arrive at the mills processes one image at a time, these studies used image mixed. It is therefore necessary to visually classify the feature extraction on a large number of images, followed logs at the scaling operation prior to sawing. While wood by multivariate data analysis. The last food classification species classification is carried out manually, the length study utilised grey-level co-occurrence matrices (GLCM) and diameter of logs are assessed automatically. Cur- (Haralick et al. 1973), a method used for feature extrac- rently there are no techniques available for species clas- tion from textures. The resulting co-occurrence matrix sification of roundwood. There are, however, several from an 8-bit greyscale image will be a square, i=j, where techniques for automated sorting and grading of logs each side represents 28s256 intensities. Each element 2 D. Nilsson and U. Edlund

of the matrix gives the probability that two pixels exist Experimental with intensity given by the two axes at a certain distance, d. The distance is preferably set to a value that corre- Sampling sponds to the local structural variations seen in the images. The distance can be measured at several angles, but A total of 79 wood discs of Norway spruce and 67 discs of Scots pine were collected from a clearing in the Va¨ sterbotten region can be set to a direction that corresponds to the image of Sweden. The clearing was made in a mature tree stand, where texture variation. GLCM calculation gives detailed infor- the age of the trees was 80 years. The 146 discs, with a thick- mation about the grey-level variation in the image. Visua- ness of approximately 50 mm, were cut from the thicker end of lisation of a GLCM is a square-shaped symmetric image. each log, and several discs could be collected from each unique If GLCM variation is mainly evident close to the diagonal, tree. Due to practical reasons, it was therefore impossible to this means that the distance is wrongly set and corre- determine the actual number of trees studied. The diameter of sponds poorly to the texture coarseness. the discs was between 95 and 245 mm for the spruce samples, The fast-Fourier transform (FFT) (Gonzalez and Woods and between 117 and 265 mm for the pine samples. In general, 2002) is a widely used method that can be applied to an the discs collected were fairly representative of the thinner image to transfer information to the frequency domain. dimensions of roundwood processed by Swedish sawmills. The bark of spruce is quite homogenous compared to that of pine, The use of a two-dimensional discrete FFT (2D DFT) has which can be of a rougher nature at the bottom of the tree. recently been used in image analysis (Wettimuny and Nevertheless, this variation was covered within the sample Penumadu 2004). Moreover, to capture both space and material, as 23 of the 67 pine discs had the rougher type of bark. frequency information, a wavelet-based transformation (Gonzalez and Woods 2002) can be used. Wavelet tex- Digital image acquisition and processing ture analysis (WTA) has been used for feature extraction in digital imaging (Van de Woewer et al. 1999; Antonelli All wood discs were partly debarked and two images (size et al. 2004). Here, the two-dimensional discrete wavelet 2048=3072 pixels) were acquired from each disc; one image transform (2D DWT) is used to analyse images at multiple with the bark left intact and one image with the bark partly resolutions. Wavelet functions are used to calculate the peeled off. The bark was manually peeled using a two-handed split-knife. All images also contained part of a reference grey differences between the resolutions or so-called scales. card (RGB 128, 128, 128). It should be noted that the discs did Wavelet transformation also means that processed imag- not cover the whole image frame. An image labelled ‘‘A’’ of es are compressed. Optimisation of the wavelet com- 300=600 pixels (equivalent to 22 mm=44 mm) containing bark position while preserving image features has also been was captured from the first image. An image of 300=600 pixels investigated (Lo et al. 2003). labelled ‘‘B’’ showing a mix of bark and wood was captured from Multivariate techniques such as principal component the second image. An image of 200=300 pixels labelled ‘‘C’’ analysis (PCA) (Wold et al. 1987) and partial least-squares containing pure wood was also captured from the second projections to latent structures (PLS) (Geladi and Kowal- image. It should be noted that since the ‘‘B’’ and ‘‘C’’ images ski 1986) are then frequently used for characterisation were captured from the same original image, the smaller size of and classification of multidimensional image data. While the ‘‘C’’ images was to avoid these cropped areas coinciding with the ‘‘B’’ images. This could, of course, be avoided by peel- these methods are directly applicable to multivariate ing the bark from a larger area. However, the size of the images images, it is often better to extract image-texture features did not impact on subsequent analysis. Figure 1a–c gives a visu- prior to multivariate modelling (Bharati et al. 2004). al overview of the captured images. The images were acquired with a tripod-mounted Canon EOS Digital Rebel 6.3-megapixel Objective of the study single-lens reflex camera using a Canon 18–55-mm lens. The following camera settings were used during image acquisition: The current study was intended as a pilot study to investigate the possibility of using multivariate image analysis as an automated method for the classification of softwood species of pine and spruce. The use of an automated method for species classification might replace the manual assessment of logs, which would facilitate the sawing process and thus be beneficial from an economic perspective. The proposed method is based on image acquisition through a digital camera coupled with multivariate image analysis. Spatial information of the image textures is extracted by 2D DWT. The wavelet variables containing both global space and frequency information are then unaligned by 2D DFT to the frequency domain. The wavelet transform merely serves as a compression tool that considerably reduces the number of variables, while conserving pixel variation found in the image textures. Furthermore, the use of grey-level co-occurrence matrices was also tested for the extrac- Figure 1 The four types of images used in the present study. tion of further information from the bark texture. PCA and (a) ‘‘A’’ image containing bark; (b) ‘‘B’’ image with bark partly PLS are then utilised for species classification of the peeled off; (c) ‘‘C’’ image containing wood only; and (d) visuali- images. sed co-occurrence matrix. Image analysis for species classification 3

ISO sensitivity, 100; aperture, f/16; shutter speed, 1/30 s; focal number of observations. It also meant that the images were ana- length, 55 mm; distance from object to lens, 500 mm; acquired lysed in more detail and that several predictions could be made image type, RAW. The built-in flash of the camera was used. All for each original image. image acquisition used the same exposure, determined by exposing the reference grey card using the flash exposure lock Multivariate classification of DWT- and DFT- function of the camera. The RAW images were then converted processed images to 24-bit TIFF format. Principal component analysis (Wold et al. 1987) was used to Texture analysis compress the wavelet coefficient matrices into score and loading vectors. PCA is a powerful method for reducing a multi- Two different approaches were taken to analyse the images: dimensional space into a few variables, or so-called principal components. Each of the compressed matrices was expressed 1. 2D DWT followed by 2D DFT; and as: 2. GLCM calculations followed by 2D DWT. s q s q q q The 2D discrete wavelet transform was used to transform raw X TP9 E tp119 tp 229 « E (1) RGB bark and wood images and the corresponding grey-level co-occurrence matrices. Simca CoDec 1.02 software (Umetrics where X is the original matrix, T and P9 are the scores and load- AB, Sweden) was used for the transformations, for which the ings matrices, and E is the error matrix. PCA was mainly per- Symmlet 8 wavelet basis function (Daubechies 1992) was utili- formed to provide visualisation for the classification of pine and sed. The length or a height of an image sized N, where N is a spruce observations. This was accomplished by plotting the q power of 2, can be represented by log2N 1 different resolutions score vectors, e.g., t1 and t2, against each other. Briefly, the or wavelet scales. The lowest resolution is represented by 1=1 score plots a visualisation of how observations are located in pixels, which is merely a mean value of the actual image. This relation to each other in the reduced variable space. Observa- value will also serve as an offset, i.e., by how much a certain tions of similar properties are located close to each other. In image deviates in that particular colour channel. As the actual addition, loading plots were used to gain understanding about images were sized 300=600 or 200=300 pixels, they were pad- the importance of the variables or, here, wavelet coefficients. In = = ded with zeros to sizes 512 1024 or 256 512 pixels to achieve this type of plot, the loading vectors, e.g., p1 and p2, are plotted dyadic compliance. No padding was necessary for the GLCM against each other. Variables of high importance are located far data, as they were sized 256=256 pixels. For RGB images, each from the coordinate origin, while variables of low importance are of the three colour channels was treated as one image. The located close to it. images were compressed into 5000 wavelet coefficients in each To further classify the observations, PLS (Geladi and Kowalski colour channel, which equals a compression ratio close to 100. 1986) was used to create prediction models between the image The maximum number of wavelet coefficients that could be data matrices and the wood species class. PLS is a commonly obtained from the RGB images was 512=1024s219s524 288. used method to find covariance between two data matrices. Use of GLCM data was the second approach in the image Instead of using PLS, other techniques such as artificial neural analysis process. The pixel distance, d, was set to 20, as much networks (Jain et al. 1996) or principal component regression of the texture variation was rather fine. The angle, u, was set to (Geladi and Esbensen 1991) might be suggested. However, PLS the horizontal direction, as this corresponded well to the bark was chosen due to its ability to model the relationships studied structure. Figure 1d shows a visualised GLCM. Matlab 6.5 between X and Y in low-rank models with high interpretability. (Mathworks) was used for GLCM calculation. Each of the PLS models (I–VI) was written in the form: The different resolutions, i.e., wavelet scales, of every image were Fourier-transformed using 2D DFT, with Fourier transfor- ysXbqf (2) mation implying no change in matrix size. All images except the GLCM data were Fourier-transformed, as these data contained where y is the wood species class, set to ‘‘1’’ for pine and ‘‘2’’ no global image space information. for spruce, X is the original matrix, b the regression coefficient and f the error. For the last model (VII), two dummy variables Generating calibration and test sets were used as a Y matrix to discriminate between the two classes. The use of two dummy variables was required to calculate Prior to multivariate data analysis, the wavelet coefficients of a mean prediction value for each original ‘‘B’’ image. Generally, each image were arranged as one vector, i.e., one observation. separate dummy variables that denote class should always be The observations were assembled in matrices, which were divid- used. A single variable can be used, however, for two classes ed into calibration sets (CS) and test sets (TS). The data matrices without introducing incorrect proportionality between the two were used to create the following models: classes. As described earlier, observations of the modelled matrices were divided into calibration sets that consisted of two- s s Model I: CS 98 ‘‘A’’ images, TS 48 ‘‘A’’ images, 5000 variables thirds of the observations and test sets that comprised the in each colour channel; remaining one-third. Centring of both X and y variables was used s s Model II: CS 98 ‘‘B’’ images, TS 48 ‘‘B’’ images, 5000 as data pre-treatment for all PLS and for the previously variables in each colour channel; described PCA models. The number of significant PLS compo- s s Model III: CS 98 ‘‘C’’ images, TS 48 ‘‘C’’ images, 5000 nents was determined by cross-validation, along with its related variables in each colour channel; Q2Y value (Wold 1978). A Q2Y value close to unity means that s s Model IV: CS 98 ‘‘A’’ GLCM, TS 48 ‘‘A’’ GLCM, 5000 the model has optimal prediction properties, while a value lower variables; than or equal to zero implies that the model has no prediction s s Model V: CS 98 ‘‘B’’ CLCM, TS 48 ‘‘B’’ GLCM, 5000 variables; ability whatsoever. The test sets were then validated externally s s Model VI: CS 98 ‘‘C’’ GLCM, TS 48 ‘‘C’’ GLCM, 5000 and the cut-off between the species was set to 1.5 for the six variables; and first models. For the seventh model, the highest value, i.e., )0.5, s s Model VII: CS 96 ‘‘A’’ GLCM, TS 50 ‘‘B’’ GLCM, 5000 denoted the species class. Observations that lay on the wrong variables. side of the given cut-off values were considered to be wrongly The last model (VII) was calculated from sub-images of the classified. Simca Pq 10.5 software (Umetrics AB) was used for original ‘‘A’’ images. The division into sub-images gave a higher calculation of all the multivariate models. 4 D. Nilsson and U. Edlund

Figure 2 Scores for the first two components, tw1x and tw2x, of the PCA model calculated from wavelet and Fourier-transformed ‘‘A’’ images containing bark. There is strong separation between the spruce (%) and pine observations (s).

Results and discussion lost almost all of its prediction efficiency when the wavelet coefficients that represented colour offset were Classification using DWT and DFT on ‘‘A’’, ‘‘B’’ and removed. Table 1 gives an overview of the seven PLS ‘‘C’’ images (models I–III) models constructed in this study. PLS models were also constructed for the ‘‘B’’ and ‘‘C’’ images. These classified The best separation of ‘‘A’’ images was observed in the the observations with an accuracy of 81.3% and 66.7%, score plot for the first and second component (Figure 2) respectively. Furthermore, both these models showed for a PCA model calculated from Fourier-transformed that the most important variables were the wavelet coef- wavelet scales. The PCA loading plot for the first and ficients accounting for colour offset. second component (Figure 3) showed that the most important wavelet coefficients were those denoting the Classification using DWT of GLCM data (models offset in each of the three colour channels, i.e., deviation IV–VI) from the mean value in the particular channel. A two- component PLS model with a Q2Y value of 0.64 was cal- The score plot for the first two components (Figure 4) of culated and used for predicting the test set observations. the PCA calculated for the ‘‘A’’ images shows obvious Five observations were wrongly predicted, which corre- separation between the pine and spruce objects. The sponds to a prediction accuracy of 89.6%. This model loading plot for the same components showed that many

Figure 3 Loadings for the first two components, pw1x and pw2x, of the PCA model calculated from the DWT and DFT processed ‘‘A’’ images containing bark shows that the wavelet offset coefficients (marked with arrows) mainly contribute to the separation between pine and spruce. Image analysis for species classification 5

Table 1 Overview of the seven PLS models used for roundwood classification.

PLS Data type NR2YQ2Y Correctly model classified (%)

I DWT ‘‘A’’ 2 0.75 0.64 89.6 II DWT ‘‘B’’ 3 0.68 0.52 81.3 III DWT ‘‘C’’ 3 0.28 0.12 66.7 IV Cooc ‘‘A’’ 6 0.99 0.92 97.9 V Cooc ‘‘B’’ 6 0.98 0.88 85.4 VI Cooc ‘‘C’’ 8 0.98 0.69 72.9 VII Cooc ‘‘A’’q‘‘B’’ 7 0.91 0.83 92.0 N, number of significant PLS components; R2Y, the explained variation in y or Y; Q2Y, the value given by cross-validation. coefficients contributed to the separation. The PLS mod- PLS model had seven components and a Q2Y value of el used for the prediction of the externally validated ‘‘A’’ 0.83, and managed to predict 46 out of 50 observations images had a Q2Y value of 0.92. Only one out of 48 correctly, which corresponds to 92%. Inarguably, many observations was wrongly predicted, which corresponds of the predicted sub-images contained both wood and to a prediction accuracy of 97.9%. Prediction models bark, so the number of sub-images should preferably be were also calculated for the ‘‘B’’ and ‘‘C’’ observations. greater. Wavelet transformation of GLCM data from They were able to classify the external test sets with an numerous sub-images can be rather time-consuming accuracy of 85.4% and 72.9%, respectively. (i.e., )1 h) if thousands of observations are handled at the same time. However, wavelet transformation of many Classification using DWT of GLCM data from sub- images can be performed in several steps, which makes images (model VII) it possible to construct large calibrations sets. The calculation times for unknown observations that are pre- All ‘‘A’’, ‘‘B’’ and ‘‘C’’ images were divided into six sub- dicted should be of no concern, even if numerous sub- images each prior to calculating a PCA model from images are predicted from one original image. GLCM data. The score plot for the first two components (Figure 5) of the PCA calculated for all sub-images shows that observations containing bark (‘‘A’’) are easily distin- Discussion guishable from observations containing wood (‘‘C’’). The observations containing both bark and wood (‘‘B’’) are The present study has shown that multivariate image spread among the two classes. A calibration set solely analysis can be used to accurately separate the species consisting of sub-images from the ‘‘A’’ images was con- spruce and pine. With further developments, the meth- structed. A PLS model was then calculated and used for odology presented here could be used in an automated the prediction of those ‘‘B’’ sub-images that were clas- sorting system for pine and spruce. In addition to instru- sified as bark according to the previous PCA model. mentation set-up, software has to be written that first Thus, several predictions were averaged into one mean divides the acquired images into proper sub-images if prediction for each original ‘‘B’’ image. The constructed needed. Secondly, suitable image extraction techniques

Figure 4 Scores for the first two components, tw1x and tw2x, of the PCA model derived from the co-occurrence matrices calculated from the ‘‘A’’ images containing bark. The spruce (%) and pine observations (s) form two classes. 6 D. Nilsson and U. Edlund

Figure 5 Scores for the first two components, tw1x and tw2x, of the PCA model from 1744 bark ‘‘A’’, barkqwood ‘‘B’’ and wood ‘‘C’’ sub-images. The bark (j) is easily separated from the wood (q). The mixed ‘‘B’’ images (s) can either be bark or wood. have to be applied, e.g., GLCM and wavelet transfor- on the wavelet scale representing the mean value of each mation; and thirdly, multivariate prediction models have colour channel. Evidently, the texture of bark exhibits too to be constructed. The images acquired should prefera- much variation, which makes it difficult to classify the two bly be representative of roundwood processed by the wood species without further image processing. How- actual sawmill and they should be obtained on site. It ever, the use of GLCM in conjunction with wavelet trans- should be noted that the images processed in the pres- formation shows that it is possible to classify pine and ent study were acquired under laboratory conditions and spruce using the bark texture. This classification is al- real images acquired on site could differ from these. most error-free for images containing bark only. Roundwood transportation and handling could cause a Furthermore, it is necessary to determine the number type of debarking that is more inhomogeneous compared of significant PLS components of a possible prediction to manual debarking, which almost completely removed model. In the present study, the number of significant the bark. Once the software is written, it could be pos- components varied between two and eight. Essentially, sible to construct unique models adapted for certain a PLS model with a single y vector should only consist conditions and that take the differences in debarking into of one significant component. PLS modelling in general account. To improve species classification, it might also tends to give models of higher rank due to orthogonal be possible to perform the prediction in several steps, variation in the X data. The first three models of the pres- whereby doubtfully classified observations could be ent study were of lower rank, mainly due to the fact that reclassified, i.e., an extra image could be acquired from three X variables accounted for much of the modelled the analysed log. In this case a higher cut-off limit could variation. The higher-rank PLS models were constructed be used to discriminate uncertain observations. Instead from GLCM data. The wavelet coefficients obtained from of using sub-images, it might also be possible to use these data are not easily modelled in a few significant segmentation techniques (Gonzalez and Woods 2002) for components without removing orthogonal X variation. separating wood and bark in acquired images. Therefore, data pre-treatment using orthogonal signal Compared to the other techniques mentioned in the correction (OSC) (Wold et al. 1998) or orthogonal PLS (O- Introduction (X-ray scanning, FT-NIR and Raman spec- PLS) (Trygg and Wold 2002) might be suggested. This troscopy, ultrasonic signalling), a possible image analysis approach was also tested, but results showed that each set-up using a digital camera would provide a convenient orthogonal component removed accounted for little var- solution for an automated wood-species classification iation in X and a large number of components was need- system. This is valid in both practical and economic ed to reduce X to variation that could be used to terms. Digital image acquisition is a fast method, which construct a low-rank PLS model. This shows that a large makes it possible to acquire and process several images number of PLS components are needed and there is an per second provided the computer hardware is fast apparent danger of under-fitting the models if only a few enough. Digital imaging is also a robust method with high components are used. It was necessary to nearly exhaust reproducibility. Unlike spectroscopic methods such as the variation in the y vector(s), as some models had R2Y Raman and FT-NIR (Thygesen and Lundqvist 2000), it is values approaching unity (Table 1). unaffected by changes in temperature and humidity. Changes in ambient light could easily be handled by Acknowledgements placing a reference in the line of sight of the camera. Grants from the Swedish Agriculture and Forestry Research Texture analysis based on the wavelet and Fourier Council are acknowledged. The authors are grateful for the transforms showed that classification depended solely assistance provided by Peter Christoffersson at Holmen Skog. Image analysis for species classification 7

References Jordan, R., Feeney, F., Nesbitt, N., Evertsen, J.A. (1998) Classi- fication of wood species by neural network analysis of ultrasonic signals. Ultrasonics 36:219–222. Antonelli, A., Cocchi, M., Fava, P., Foca, G., Franchini, C.G., Ulri- Kvaal, K., Wold, J.P., Indahl, U.G., Baardseth, P., Næs, T. (1998) ci, A. (2004) Automated evaluation of food colour by means Multivariate feature extraction from textural images of bread. of multivariate image analysis coupled to a wavelet-based Chemomet. Intell. Lab. Syst. 42:141–158. classification algorithm. Anal. Chim. Acta 515:3–13. Lawrence, A.H. (1989) Rapid characterization of wood species Basset, O., Buquet, B., Abouelkaram, S., Delachartre, P., Culioli, by ion mobility spectrometry. J. Pulp Pap. Sci. 15(5): J. (2000) Application of texture image analysis for the clas- J196–J199. sification of bovine meat. Food Chem. 69:437–445. Lewis, I.R., Daniel, N.W., Chaffin, N., Griffiths, P. (1994) Raman Bharati, M.H., Liu, J.J., MacGregor J.F. (2004) Image texture spectrometry and neural networks for the classification of analysis: methods and comparisons. Chemomet. Intell. Lab. wood types 1. Spectrochim. Acta A Mol. Spectrosc. 50: Syst. 72:57–71. 1943–1958. Brunner, M., Eugster, R., Trenka, E., Bergamin-Strotz, L. (1996) Lo, S.C.B., Li, H., Freedman, M.T. (2003) Optimization of wavelet FT-NIR spectroscopy and wood identification. Holzforschung decomposition for image compression and feature preser- vation. IEEE Trans. Med. Imaging 22:1141–1151. 50:130–134. Ma¨ enpa¨a¨ , T., Pietikainen, M. (2004) Classification with color Daubechies, I. Ten Lectures on Wavelets. SIAM, Philadelphia, and texture: jointly or separately? Pattern Recogn. 37: 1992. 1629–1640. Esbensen, K., Geladi, P. (1989) Strategy of multivariate image- Oja, J., Wallba¨ cks, L., Grundberg, S., Ha¨ gerdal, E., Gro¨ nlund, A. analysis (MIA). Chemomet. Intell. Lab. Syst. 7:67–86. (2003) Automatic grading of Scots pine (Pinus sylvestris L.) Funck, J.W., Zhong, Y., Butler, D.A., Brunner, C.C., Forrer, J.B. sawlogs using an industrial X-ray log scanner. Comput. Elec- (2003). Image segmentation algorithms applied to wood tron. Agric. 41:63–75. defect detection. Comput. Electron. Agric. 41:157–179. Thygesen, L.G., Lundqvist, S.O. (2000) NIR measurement of Geladi, P., Esbensen, K. (1991) Regression on multivariate imag- moisture content in wood under unstable temperature con- es – Principal component regression for modeling, prediction ditions. Part 1. Thermal effects in near infrared spectra of and visual diagnostic tools. J. Chemomet. 5:97–111. wood. J. Near Infrared Spectrosc. 8:183–189. Geladi P., Kowalski, B.R. (1986) Partial least squares regression: Trygg, J., Wold, S. (2002) Orthogonal projections to latent struc- A tutorial. Anal. Chim. Acta 185:1–17. tures, O-PLS. J. Chemomet. 16:119–128. Geladi, P., Isaksson, H., Lindqvist, L., Wold, S., Esbensen, K. Van de Wouwer, G., Scheunders, P., Livens, S., Van Dyck, D. (1989) Principal component analysis of multivariate images. (1999) Wavelet correlation signatures for color texture char- Chemomet. Intell. Lab. Syst. 5:209–220. acterization. Pattern Recogn. 32:443–451. Gjerdrum, P., Warensjo¨ , M., Nylinder, M. (2001) Classification of Wettimuny, R., Penumadu, D. (2004) Application of Fourier ana- crook types for unbarked Norway spruce sawlogs by means lysis to digital imaging for particle shape analysis. J. Comp. of a 3D log scanner. Holz Roh Werkstoff 59:374–379. Civil Eng. 18:2–9. Gonzalez R.C., Woods R.E. Digital Image Processing, 2nd ed. Wold, S. (1978) Cross-validatory estimation of the number of Prentice Hall, New Jersey, 2002. components in factor and principal components models. Hagman, P.O.G. (1993) Automatic quality sorting of Picea abies Technometrics 20:397–405. logs with a gamma-ray log scanner. Scand. J. For. Res. 8: Wold S., Esbensen, K., Geladi, P. (1987) Principal component 583–590. analysis. Chemomet. Intell. Lab. Syst. 2:37–52. ¨ Haralick, R.M., Shanmugam, K., Dinstein, I. (1973) Textural fea- Wold, S., Antti, H., Lindgren, F., Ohman, J. (1998) Orthogonal signal correction of near-infrared spectra. Chemomet. Intell. tures for image classification. IEEE Trans. Syst. Man Cyber- Lab. Syst. 44:175–185. netics 3:610–621. Jain, A.K., Mao, J.C., Mohiuddin, K.M. (1996) Artificial neural networks: A tutorial. Computer 29:31–44. Received March 2, 2005. Accepted July 8, 2005.

PEER-REVIEWED Experimental

David Nilsson1, Ulf Edlund1, Michael Sjöström1, Roland Agnemo2 The wood material used in this study consisted of 60 spruce logs which were stored in climate chambers for 14 weeks. The storage was performed according to an experimental Prediction of thermo mechanical pulp brightness using design with five variables that were assumed greatly influence the brightness of the stored NIR spectroscopy on wood raw material wood. The design had 15 experimental runs (each consisting of four logs) that were sampled on three occasions. The actual number Keywords: of interest. Pre-treatment /5/ methods are chemical changes occur in the wood and a of samples was therefore 45, and was com- Near-infrared, spectroscopy, principal com- applied to NIR spectra in order to remove study by Borga et. al. /29/ showed that it is prised of 180 wood chunks and 180 wood ponent analysis, partial least squares regres- unwanted variation. Multiplicative Scatter possible to use NIR spectroscopy to moni- discs. The chunks were later refined into sion to latent variables, thermo mechanical Correction /4/ (MSC) is a method used for tor these variations. They investigated the TMP while the discs were measured with pulp, ISO brightness correcting NIR spectra from scatter effects correlation between near-infrared spectra of near-infrared reflectance spectroscopy. such as baseline shifts and different slopes. wet-stored pine timber and the storage time The TMP was bleached with either hy- Abstract Orthogonal Signal Correction /6, 7, 8/, in relation to the water quality. It was found drogen peroxide or sodium dithionite fol- (OSC), deducts variation in the data that that chemical changes in the wood during lowed by ISO brightness measurements. See Partial Least Squares regression to latent is not correlated to the response variable wet-storage did occur and these changes Appendix 1 for full details of the bleaching structures was used to find a correlation of interest. were also related to the bacterial damage procedure as given in earlier work /30/ by between near-infrared reflectance spectra NIR spectroscopy has previously been of the wood. The ISO brightness of pulp the authors. The reader is referred to the of samples of milled Norway Spruce (Picea used to predict the content of different spe- prior to bleaching is expected to be corre- same work /30/ for more details about the abies) and ISO brightness of bleached ther- cies in a mixture /9/ and for prediction of lated to the brightness of the bleached pulp wood storage and the refining of TMP. mo mechanical pulp. Logs of spruce were strength parameters for softwood pulps /10/. but there are more factors that contribute stored in climate chambers with different Furthermore, it was used for estimation of to the bleaching outcome. For example, the Vis-Near-infrared reflectance conditions for 14 weeks. A factorial design lignin content in wood /11/, for characteri- bleachability of pulp can fluctuate due to measurements based on five variables was constructed to sation of pulp /12/ and for examination of the previous mentioned metal ion degrada- A NIR Systems 6500 Spectrophotometer control the storage and to provide different pine needle samples /13/ and several other tion of hydrogen peroxide, varying degree was used for the reflectance measurements. A brightness properties of the stored spruce applications /14/. NIR spectroscopy is a of bacterial damage and also the reactivity fibre optic probe was used for measurements logs. Wood samples were collected during powerful technique for determination of of the used bleaching chemical. This topic directly on the wood surface and a spinning the storage and they were analysed with moisture content within different sample is comprehensively covered by Lorås /19/ cup for measuring on ground wood. This in- near-infrared reflectance spectroscopy and materials /15, 16, 17, 18/. who developed two methods for determina- strument uses a Si-based detector for meas- later refined to thermo mechanical pulp. The raw material used in the pulp and tion of bleachability of thermo mechanical uring in the wavelength region of 400 to The pulp was bleached with both hydro- papermaking industry, i.e. wood, is often pulp (TMP). It is in the scope of this work 1100 nm and a PbS detector between 1100 gen peroxide and sodium dithionite. The exposed to wet-storage which has been in- to prove that NIR spectroscopy can be used to 2500 nm. Thus, the scanned region was ISO brightness of the bleached pulp was vestigated in a number of studies /19, 20, as an alternative method for predicting the 400 to 2500 nm with each spectrum con- correlated to the near-infrared spectra of 21/. Wet-storage decreases the brightness bleaching outcome of mechanical pulp and sisting of an average of 32 scans. It is notice- the untreated samples using Partial Least of wood and the bleaching of pulp will be to give an approximate value of the ISO able that the visual region was included in Squares regression to latent structures. The less efficient and higher quantities of bleach- brightness obtained after bleaching. The the measured range as well. The resolution modelling showed that it is possible to pre- ing chemicals have to be used. The main use of a fast method for prediction of wood of the instrument was limited to every even dict the ISO brightness of bleached thermo reason for the darkening of wood during quality in terms of brightness achieved af- wavelength, which resulted in spectra that mechanical pulp from near-infrared spectra wet-storage is bark staining which caus- ter bleaching could facilitate the corrections consisted of 1050 values. of original wood raw material. es migration of tannins and other organic that have to be made for the bleaching to be Four different reflectance measurements compounds from tree bark into sapwood more efficient. The prediction could also be were performed on the different stored Introduction /19/. In addition, storage often leads to used for discarding certain wood raw mate- spruce material; wood degradation by bacteria causing an rial not suitable for the manufacturing of Method A. Measurements directly on the Near-infrared reflectance (NIR) spectrosco- increased permeability of wood /22/. It has TMP. In total, the use of NIR spectroscopy wood surface py in combination with multivariate statisti- been proven several times that metal ions to predict ISO brightness could improve the Method B. Measurements on ground cal techniques has proven to be a powerful such as manganese and iron catalyse the TMP manufacturing process in both practi- wood tool to determine wood, pulp and paper decomposition /23, 24, 25/ of hydrogen cal and economical aspects. Method C. Measurements on dried ground properties. NIR spectroscopy is a fast and peroxide, one of the most commonly used wood robust method and is therefore suitable to chemicals for bleaching of mechanical pulp. Method D. Measurements on thermo me- be implemented in various processes. Due to Additionally, metal ions can form com- chanical pulp the low-resolved nature of NIR spectrosco- plexes with organic compounds that can Please see Appendix 2. for a complete py, multivariate data analysis has to be used contribute to discolouration of wood /26/. description of the four measurements meth- to extract variation from the spectral data. The brightness of pulp is usually expressed ods. Multivariate methods such as Partial Least in terms of ISO brightness /27, 28/, which 1. David Nilsson*, Ulf Edlund, Michael Sjöström, De- Squares regression to latent structures /1, 2/ is defined as the reflectance from a sheet of partment of Chemistry, Organic Chemistry, Umeå Data treatment (PLS) and Principal Component Analysis paper made from the pulp measured at the University, SE-901 87 Umeå, Sweden The NIR spectra collected according to the 2. Roland Agnemo, Domsjö Fabriker AB, SE-891 80 /3/ (PCA) can be applied for classification wavelength 457 nm. Örnsköldsvik, Sweden four methods were assembled into four dif- of material and for predicting properties During wood storage, structural and *Corresponding author ferent matrices with 45 observations and

102 Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005

Nimetön 2 1 9/9/05 9:09:58 AM PEER-REVIEWED Experimental 1050 variables or wavelengths. The spectra Computer software Matlab 5.1 /32/ and were checked for outliers using visual in- Unscrambler 7.5 /33/ were used for building The wood material used in this study con- spection of score plots created from PCA the data sets and for pre-treatment of data. sisted of 60 spruce logs which were stored in models. Prior to multivariate modelling, Simca-P 8.0 /34/ was used for pre-treat- climate chambers for 14 weeks. The storage the NIR spectra were then pre-treated with ment of data and for computing multivari- was performed according to an experimental Multiplictative Scatter Correction, MSC ate models. Spotfire Pro 4.0 /35/ was used Prediction of thermo mechanical pulp brightness using design with five variables that were assumed /4/. This pre-treatment gave overall the for making 3D plots. greatly influence the brightness of the stored best predictions predictions for ISO bright- NIR spectroscopy on wood raw material wood. The design had 15 experimental runs ness and unless stated, all prediction models Results and discussion (each consisting of four logs) that were sam- discussed in the results section were based pled on three occasions. The actual number on MSC-treated spectra. Two other filter- The four types of measurements presented chemical changes occur in the wood and a of samples was therefore 45, and was com- ing techniques were also tested; the use in this study represent different degrees of study by Borga et. al. /29/ showed that it is prised of 180 wood chunks and 180 wood of first-derivatives and Orthogonal Signal sample treatment. It is of interest that the possible to use NIR spectroscopy to moni- discs. The chunks were later refined into Correction, OSC /6/. OSC was primarily samples are taken at an early stage in the tor these variations. They investigated the TMP while the discs were measured with used in order to facilitate model interpreta- paper production chain. This makes it pos- correlation between near-infrared spectra of near-infrared reflectance spectroscopy. tion. Variable centring was the only scaling sible to make early predictions of the end- wet-stored pine timber and the storage time The TMP was bleached with either hy- technique that was applied to the spectral product properties, i.e. in this study the ISO in relation to the water quality. It was found drogen peroxide or sodium dithionite fol- data prior to the multivariate PCA and PLS brightness value. that chemical changes in the wood during lowed by ISO brightness measurements. See modelling. The measured wood raw material origi- wet-storage did occur and these changes Appendix 1 for full details of the bleaching nated from an experimental wood storage were also related to the bacterial damage procedure as given in earlier work /30/ by Multivariate modelling design that produced samples of different of the wood. The ISO brightness of pulp the authors. The reader is referred to the Principal Component Analysis was per- properties regarding bleachability and ISO prior to bleaching is expected to be corre- same work /30/ for more details about the formed on the NIR spectra. The PCA scores brightness values obtainable after bleaching. lated to the brightness of the bleached pulp wood storage and the refining of TMP. and loadings were examined to find possible The use of experimental design implies that but there are more factors that contribute clusters and important wavelength regions. the prediction models for ISO brightness to the bleaching outcome. For example, the Vis-Near-infrared reflectance PCA was also used for outlier detection. are more extensive than if the measured and bleachability of pulp can fluctuate due to measurements PLS was then utilised to find correlation bleached samples were picked randomly. the previous mentioned metal ion degrada- A NIR Systems 6500 Spectrophotometer between NIR spectra and the ISO bright- tion of hydrogen peroxide, varying degree was used for the reflectance measurements. A ness of TMP. For validation, the objects were Reference value checking of bacterial damage and also the reactivity fibre optic probe was used for measurements divided into two training sets and two test. The reference values in this study consist- of the used bleaching chemical. This topic directly on the wood surface and a spinning The PLS models were validated externally by ed of ISO brightness values for peroxide is comprehensively covered by Lorås /19/ cup for measuring on ground wood. This in- prediction of the excluded objects, i.e. the bleached and dithionite bleached TMP. In who developed two methods for determina- strument uses a Si-based detector for meas- objects in the test sets. RMSEP, Root Mean total 90 values and thus 45 values for each tion of bleachability of thermo mechanical uring in the wavelength region of 400 to Squared Error of Prediction, was calculated bleaching approach were collected. The 45 pulp (TMP). It is in the scope of this work 1100 nm and a PbS detector between 1100 in order to estimate the predictive ability of samples were comprised of 15 logs that were to prove that NIR spectroscopy can be used to 2500 nm. Thus, the scanned region was the models. The number of PLS compo- sampled on three occasions. It was possible as an alternative method for predicting the 400 to 2500 nm with each spectrum con- nents was determined by internal cross vali- to detect abnormal ISO brightness values bleaching outcome of mechanical pulp and sisting of an average of 32 scans. It is notice- dation /31/ (Q2Y) and by visual inspection that were caused by deviating storage-time to give an approximate value of the ISO able that the visual region was included in of loading plots. dependencies by plotting the 45 samples brightness obtained after bleaching. The the measured range as well. The resolution The following formula was used for cal- for each of the two bleaching approaches use of a fast method for prediction of wood of the instrument was limited to every even culation of RMSEP; with the samples grouped together three quality in terms of brightness achieved af- wavelength, which resulted in spectra that � and three. It was also possible to compare � ter bleaching could facilitate the corrections consisted of 1050 values. �� the values with the accessible ISO bright- that have to be made for the bleaching to be Four different reflectance measurements �� ness of the corresponding unbleached pulp. more efficient. The prediction could also be were performed on the different stored � (1) The ISO brightness values for the dithionite used for discarding certain wood raw mate- spruce material; where yipred denotes the ith predicted bleached TMP (Fig. 1) reveal that there rial not suitable for the manufacturing of Method A. Measurements directly on the value, yiobs the ith observed value and N the seem to be three clearly suspicious values, TMP. In total, the use of NIR spectroscopy wood surface total number of objects in the predicted the six-week stored sample in S4, the fresh to predict ISO brightness could improve the Method B. Measurements on ground test sets. Since two test sets were excluded sample in S8 and the six-week stored sample TMP manufacturing process in both practi- wood from each of the PLS model computations, in S15. These three values do not follow the cal and economical aspects. Method C. Measurements on dried ground the calculated RMSEP values were derived general trend of decreased brightness with wood from the total number of objects in the two increased storage time. To further investigate Method D. Measurements on thermo me- test sets instead of taking the mean RMSEP the outliers, these values were predicted as chanical pulp value for the test sets. Thus, the use of two an excluded test set in a PLS model between Please see Appendix 2. for a complete externally validated test sets implied that the NIR spectra obtained from Method A description of the four measurements meth- two PLS models had to be calculated for and the corresponding ISO brightness val- ods. each of the data pre-treatment methods. ues. Only differentiation was used as data 1. David Nilsson*, Ulf Edlund, Michael Sjöström, De- The number of PLS components was always pre-treatment in the model that had two partment of Chemistry, Organic Chemistry, Umeå Data treatment the same for both test sets. It should also significant components with a R2Y of 0.52 University, SE-901 87 Umeå, Sweden The NIR spectra collected according to the be noted that the test set objects were left and a Q2Y of 0.439. The observed values for 2. Roland Agnemo, Domsjö Fabriker AB, SE-891 80 Örnsköldsvik, Sweden four methods were assembled into four dif- out of the parameter estimation for MSC these three objects are clearly lower than ex- *Corresponding author ferent matrices with 45 observations and and OSC. pected (Fig .2). Furthermore, a PCA model ☛

Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005 103

Nimetön 2 2 9/9/05 9:10:00 AM Fig. 4. Prediction of test sets excluded from PLS modelling between ISO brightness of dithionite bleached samples and spectra measured on wood surface.

Fig. 3. Score plot for the ﬁrst three components of the PCA model calculated from spectra measured on wood surface. The darker objects were subjected to low levels of watering during storage.

that were exposed to a high irrigation level correspond very well to those objects that Fig. 1. ISO brightness values for dithionite bleached samples. The were clearly influenced by bark staining ac- three values that originated from one main sample were plotted cording to a visual inspection. Furthermore, Fig. 5. Loading plot for the first component together in order to find outliers. two objects in the top left corner are slight- of the PCA model calculated from spectra ly separated from the rest of the objects. measured on ground wood. The first loading values for the peroxide bleached pulps were A common feature of these two objects is was heavily influenced by variation in water rejected as well. that they exhibited a perceptible yellowing content. on the wood surface. Overall, the three-di- Method A, Multivariate modelling of mensional space described by the first three Table 1. Summary of all PLS models made for prediction of dithionit NIR spectra measured on wood surface principal components gives an overview of bleached TMP brightness. PCA was used to calculate a model from differences between the wood samples that is Method1 Pre-Treatment2 A3 Q2Y TS-14 Q2Y TS-25 RMSEP6 centred and multiplicative scatter corrected due to various storage conditions. Based on A 1:st Deri. 2 0.515 0.467 1.036 NIR spectra. The model had three signifi- the score plot for the first three components, MSC 2 0.605 0.592 0.958 cant components that explained 98.5% of two test sets that covered most of the sample OSC {1} 1 0.771 0.719 0.961 2 the data with respect to R X and 98.1% variation were selected and used for exter- MSC, Visual 3 0.533 0.516 0.976 2 with respect to Q X. The score plot for the nal validation of the models predicting ISO B 1:st Deri. 4 0.261 0.365 1.277 first two components (Fig. 3) shows that brightness of dithionite bleached TMP. The MSC 2 0.348 0.327 1.281 the wood samples were clustered into two same test sets were used for all data sets in OSC {1} 1 0.387 0.709 1.138 groups as labelled in the plot. One group order to allow for fair comparisons between MSC, Visual 3 0.194 0.406 1.297 Fig. 2. Observed vs. Predicted plot for is comprised of samples that were exposed the different sample treatments. C 1:st Deri. 6 0.759 0.701 1.024 dithionite bleached samples. All predicted to a high irrigation level during the wood The two-component PLS models made MSC 3 0.492 0.520 1.132 objects were included in the model that was storage. The other group consists of sam- for the prediction of the dithionite bleached OSC {1} 1 0.838 0.888 1.240 comprised on spectra measured on wood ples that were subjected to a low irrigation samples, revealed a correlation between the MSC, Visual 3 0.402 0.460 1.275 surface. level. The corresponding loading plot for the NIR spectra and the ISO brightness with MSC, NIR 3 0.370 0.347 1.119 first three components showed that the first a RMSEP value of 0.96 and Q2Y values of D 1:st Deri. 3 0.771 0.785 0.809 and third components primarily describe 0.61 and 0.59. The RMSEP value should be MSC 2 0.822 0.822 0.714 was created using the brightness values as spectral variation in the visual region. The compared to the standard deviation (1.49) OSC {1} 1 0.836 0.873 0.962 X-block. The distance to model plot for second component is comprised of the vari- for the observed ISO brightness values of MSC, Visual 3 0.811 0.818 0.754 the first component showed that the three ation in the visual region, but also contains the test set objects in order to get an esti- 1. The used sample pre-treatment method. suspicious values had the largest deviation variation that is due to hydrogen – oxygen mation of how well the predictions can be 2. The used data pre-treatment technique. The bracketed digit denotes the among all samples. Therefore, these values deformation and stretching around the made in the given interval. An observed number of removed orthogonal components. 3. The number of used PLS components. were excluded from further modelling of wavelength of 1950 nm, as well as varia- vs. predicted plot for the prediction of the 4. The cross-validated Q2Y-value of the first test set. dithionite bleaching prediction. In the same tion in the higher wavelengths of 2100 to dithionite bleached test set objects can be 5. The cross-validated Q2Y-value of the second test set. way as described above, three ISO brightness 2300 nm. It is noticeable that the objects seen in Fig. 4. 6. The Root Mean Squared Error of Prediction value.

104 Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005

Nimetön 2 3 9/9/05 9:10:13 AM The predictions made for the peroxide in the sample material although it is much bleached samples were, however, less suc- less prominent. The ﬁrst component mainly cessful than those for the dithionite bleach- contains variation that is due to water con- ing. The PLS modelling found very little tent and this can be seen in the correspond- correlation between the X and y data even ing loading plot (Fig. 5) The loading plot of when using as many as ten components. the second component (not shown) covered This was also reﬂected in the RMSEP val- the variation found in the visual region of ue (1.32) and in the Q2Y values (~0). The the spectra. RMSEP value is higher than the standard The PLS models calculated for the deviation of the test set objects (1.18) which dithionite bleached samples revealed a mod- proves that the models have no predictive erate correlation that had a RMSEP value of ability. Hereafter, the Q2Y values will not 1.28. Compared to the standard deviation be reported in the text and the number of for the observed values in the test sets, the PLS components will be presented more RMSEP value is just barely lower than the Fig. 4. Prediction of test sets excluded from sparsely. The reader is referred to Table 1 standard deviation for the observed values PLS modelling between ISO brightness of and for a complete summary of the charac- in the both test sets. ☛ dithionite bleached samples and spectra teristics of all prediction models calculated measured on wood surface. in this study.

Method B, Multivariate modelling of NIR spectra measured on ground wood A PCA model with five components ex- plaining nearly 100% of the data was calcu- Fig. 3. Score plot for the first three components of the PCA model lated from MSC-treated spectra in order to calculated from spectra measured on wood surface. The darker get an overview of the milled samples. The objects were subjected to low levels of watering during storage. score plot for the first three components (not shown) revealed that there is some separa- that were exposed to a high irrigation level tion between different levels of irrigation. correspond very well to those objects that This grouping is however much less obvi- were clearly influenced by bark staining ac- ous when compared to the object grouping cording to a visual inspection. Furthermore, Fig. 5. Loading plot for the first component measured according to Method A. As for two objects in the top left corner are slight- of the PCA model calculated from spectra method A, the effect of bark staining as well Fig. 6. Prediction of excluded test sets ly separated from the rest of the objects. measured on ground wood. The first loading as wood yellowing could also be seen using objects for the modelling between spectra A common feature of these two objects is was heavily influenced by variation in water a different object labelling. It is obvious that of ground wood and ISO brightness of that they exhibited a perceptible yellowing content. the effect of wood storage still is significant peroxide bleached TMP. on the wood surface. Overall, the three-dimensional space described by the first three Table 1. Summary of all PLS models made for prediction of dithionit Table 2. Summary of all PLS models made for prediction of peroxide principal components gives an overview of bleached TMP brightness. bleached TMP brightness. differences between the wood samples that is Method1 Pre-Treatment2 A3 Q2Y TS-14 Q2Y TS-25 RMSEP6 Method1 Pre-Treatment2 A3 Q2Y TS-14 Q2Y TS-25 RMSEP6 due to various storage conditions. Based on A 1:st Deri. 2 0.515 0.467 1.036 A 1:st Deri. 10 0.073 0.399 1.390 the score plot for the first three components, MSC 2 0.605 0.592 0.958 MSC 10 ˜0 0.132 1.319 two test sets that covered most of the sample OSC {1} 1 0.771 0.719 0.961 OSC {2} 1 0.699 0.539 1.453 variation were selected and used for exter- MSC, Visual 3 0.533 0.516 0.976 MSC, Visual >10 ˜0 ˜0 N/A nal validation of the models predicting ISO B 1:st Deri. 4 0.261 0.365 1.277 B 1:st Deri. 3 0.517 0.313 0.955 brightness of dithionite bleached TMP. The MSC 2 0.348 0.327 1.281 MSC 5 0.668 0.429 0.788 same test sets were used for all data sets in OSC {1} 1 0.387 0.709 1.138 OSC {1} 1 0.732 0.587 0.814 order to allow for fair comparisons between MSC, Visual 3 0.194 0.406 1.297 MSC, Visual 3 0.472 0.212 0.905 the different sample treatments. C 1:st Deri. 6 0.759 0.701 1.024 C 1:st Deri. 7 0.797 0.829 0.859 The two-component PLS models made MSC 3 0.492 0.520 1.132 MSC 3 0.574 0.459 0.744 for the prediction of the dithionite bleached OSC {1} 1 0.838 0.888 1.240 OSC {1} 1 0.925 0.972 0.776 samples, revealed a correlation between the MSC, Visual 3 0.402 0.460 1.275 MSC, Visual 3 0.521 0.552 0.732 NIR spectra and the ISO brightness with MSC, NIR 3 0.370 0.347 1.119 MSC, NIR 3 0.533 0.472 0.846 a RMSEP value of 0.96 and Q2Y values of D 1:st Deri. 3 0.771 0.785 0.809 D 1:st Deri. 2 0.512 0.413 0.866 0.61 and 0.59. The RMSEP value should be MSC 2 0.822 0.822 0.714 MSC 2 0.477 0.472 0.951 compared to the standard deviation (1.49) OSC {1} 1 0.836 0.873 0.962 OSC {1} 1 0.720 0.797 1.019 for the observed ISO brightness values of MSC, Visual 3 0.811 0.818 0.754 MSC, Visual 3 0.515 0.467 0.740 the test set objects in order to get an esti- 1. The used sample pre-treatment method. 1. The used sample pre-treatment method. mation of how well the predictions can be 2. The used data pre-treatment technique. The bracketed digit denotes the 2. The used data pre-treatment technique. The bracketed digit denotes the made in the given interval. An observed number of removed orthogonal components. number of removed OSC components. 3. The number of used PLS components. 3. The number of used PLS components. vs. predicted plot for the prediction of the 4. The cross-validated Q2Y-value of the first test set. 4. The cross-validated Q2Y-value of the first test set. dithionite bleached test set objects can be 5. The cross-validated Q2Y-value of the second test set. 5. The cross-validated Q2Y-value of the second test set. seen in Fig. 4. 6. The Root Mean Squared Error of Prediction value. 6. The Root Mean Squared Error of Prediction value.

Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005 105

Nimetön 2 4 9/9/05 9:10:23 AM The prediction of ISO brightness of Interpretation of the PLS models peroxide bleached TMP based on spectra The OSC algorithm removes variation in measured on ground wood was significantly the X data that is not correlated to y by cal- improved compared to the correlation made culating a set of score and loading vectors from the wood surface measurements. The that are orthogonal to y and are used for RMSEP value calculated from the predicted filtering X. It is possible to get an under- test sets was 0.79. standing of what spectral variation that was removed by plotting the loading of an OSC Method C, Multivariate modelling of component. The orthogonal signal corrected NIR spectra measured on dried ground spectra can then be correlated to y using a wood PLS model that only consists of one signifi- The reason for drying wood before meas- cant component. The PLS loading weights uring the samples with NIR spectroscopy for that single component yields informa- was to limit the influence of varying mois- tion on regions of the spectral data that are ture content within the sample material. correlated to y. The drawbacks of treating the samples this The OSC loadings (Fig. 7a–h), denoted way could be oxidation and lignin colour p, and the PLS weights (Fig. 8 a–h), denoted changes. PCA was performed on the MSC- w, were plotted for all prediction models treated NIR matrix with a five-component based on orthogonal signal corrected spec- model that explained nearly 99% of the vari- tra. All different sample treatment methods ance of the spectral data. The score plot for were included, which resulted in that eight the first three components showed that the plots were made for each of the two bleach- sample material still is separated due to the ing procedures. It is important to remember irrigation level during storage, which main- that the size of the loadings or weights given ly can be seen in the second component. by the scale on y-axis cannot be compared The loading plot for the first three compo- directly between two different plots since nents confirmed that the second component only w is normalised. depends on variation in the wavelengths It can be seen in Fig. 7a and Fig. 7b that around 1950 nm, which is due to the ear- the variation removed prior to computing lier mentioned hydrogen-oxygen stretching the models based on spectra recorded ac- and deformation in water. The samples also cording to Method A, mainly lies in the differ with respect to variation in the visual visual region. The two peaks that are due region, which can be seen in the first and to water absorption are also removed. The third component. weight plots (Fig. 8a and Fig. 8b) show that The PLS models made for the prediction there is no variation in these plots that origi- of ISO brightness of the dithionite bleached nates from the water peaks. Consequently, pulp gave a RMSEP value of 1.02. The ISO the difference between Fig. 7a and Fig. 7b brightness of the peroxide bleached TMP was lies in the visual region and the OSC has also well modelled using spectra measured on therefore removed more variation in the vis- the milled and dried wood. The external vali- ual region prior to computing the model us- dation gave a RMSEP value of 0.74, which ing peroxide bleached ISO brightness than is significantly lower than the standard de- it has prior to computing the dithionite- viation of 1.18 calculated from the observed bleaching model. values of the test set objects. The prediction The OSC loading plots made for the of ISO brightness for these models is accept- ground wood (Fig. 7c and Fig. 7d) show able given that the interval in ISO brightness that the orthogonal component for both is rather narrow (Fig. 6.). bleaching procedures is, as above, primarily composed of variation that is due to the wa- Method D, Multivariate modelling of Fig. 7. OSC loading plots for all sample pre-treatment methods. The OSC loadings were ter content. Once again, the ISO brightness NIR spectra measured on TMP taken from the first orthogonal component in each case, and represent the variation of the dithionite-bleached TMP seems to be Performing NIR measurements on TMP orthogonal to ISO brightness better correlated to variation in the visual might not be a viable solution in practise, a) dithionite bleaching/Method A e) dithionite bleaching/Method C region (Fig. 8c) compared to the brightness since the unbleached TMP samples repre- b) peroxide bleaching/Method A f) peroxide bleaching/Method C achieved after peroxide bleaching. Peroxide sent a stage in the production process that c) dithionite bleaching/Method B g) dithionite bleaching/Method D bleached samples better correlate to changes is close to the final product. These meas- d) peroxide bleaching/Method B h) peroxide bleaching/Method D in the near-infrared region (Fig. 8d). For the urements were still included in the study in dried samples, the loading plots for the OSC order to examine how well the predictions 100% of the variation in data. Only a slight unbleached pulp samples. The validation of pre-treated spectra measured according to based on these spectra compared to the pre- grouping due to the effect of bark staining the dithionite bleached samples generated Method C (Fig. 7e and Fig. 7f) showed as dictions made from the earlier material. The was noticeable in the score plot for the first a RMSEP value of 0.71 while the peroxide expected that that the contribution from the PCA that was performed on the multiplica- three components. bleached samples were slightly worse pre- water peaks was minor. There is still some tive scatter corrected TMP spectra found a As expected, the ISO brightness was well dicted with a RMSEP value of 0.95. variation left in the PLS weights plots (Fig. five-component model that explained nearly predicted from the spectra measured on the 8e and Fig. 8f) that originates from the wa-

106 Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005

Nimetön 2 5 9/9/05 9:10:27 AM Interpretation of the PLS models The OSC algorithm removes variation in the X data that is not correlated to y by calculating a set of score and loading vectors that are orthogonal to y and are used for filtering X. It is possible to get an understanding of what spectral variation that was removed by plotting the loading of an OSC component. The orthogonal signal corrected spectra can then be correlated to y using a PLS model that only consists of one significant component. The PLS loading weights for that single component yields information on regions of the spectral data that are correlated to y. The OSC loadings (Fig. 7a–h), denoted p, and the PLS weights (Fig. 8 a–h), denoted w, were plotted for all prediction models based on orthogonal signal corrected spectra. All different sample treatment methods were included, which resulted in that eight plots were made for each of the two bleaching procedures. It is important to remember that the size of the loadings or weights given by the scale on y-axis cannot be compared directly between two different plots since only w is normalised. It can be seen in Fig. 7a and Fig. 7b that the variation removed prior to computing the models based on spectra recorded according to Method A, mainly lies in the visual region. The two peaks that are due to water absorption are also removed. The weight plots (Fig. 8a and Fig. 8b) show that there is no variation in these plots that originates from the water peaks. Consequently, the difference between Fig. 7a and Fig. 7b lies in the visual region and the OSC has therefore removed more variation in the visual region prior to computing the model using peroxide bleached ISO brightness than it has prior to computing the dithionite- bleaching model. The OSC loading plots made for the ground wood (Fig. 7c and Fig. 7d) show that the orthogonal component for both bleaching procedures is, as above, primarily composed of variation that is due to the wa- Fig. 7. OSC loading plots for all sample pre-treatment methods. The OSC loadings were ter content. Once again, the ISO brightness taken from the first orthogonal component in each case, and represent the variation of the dithionite-bleached TMP seems to be orthogonal to ISO brightness better correlated to variation in the visual Fig. 8. PLS weights plots for all sample pre-treatment methods. The weights, w, were taken e) dithionite bleaching/Method C region (Fig. 8c) compared to the brightness from the first component calculated for the correlation between ISO brightness and f) peroxide bleaching/Method C achieved after peroxide bleaching. Peroxide spectral data. g) dithionite bleaching/Method D bleached samples better correlate to changes h) peroxide bleaching/Method D in the near-infrared region (Fig. 8d). For the ter content but the peak around 1950 nm is the sample material (Fig. 7g and Fig. 7h). dried samples, the loading plots for the OSC less dominant, but sharper. The ISO bright- Still, there is a considerable amount of vari- unbleached pulp samples. The validation of pre-treated spectra measured according to ness of dithionite bleached TMP appears ation in the water-influenced regions of the the dithionite bleached samples generated Method C (Fig. 7e and Fig. 7f) showed as once again to be more affected by the visual spectra measured on TMP (Method D) that a RMSEP value of 0.71 while the peroxide expected that that the contribution from the region of the spectra. is correlated to the ISO brightness of both bleached samples were slightly worse pre- water peaks was minor. There is still some It is evident that the OSC pre-treat- peroxide- and dithionite-bleached TMP dicted with a RMSEP value of 0.95. variation left in the PLS weights plots (Fig. ment has removed much of the variation (Fig. 8g and Fig. 8h). Furthermore, it is 8e and Fig. 8f) that originates from the wa- that is due to varying water content within clearly so that the dithionite bleached TMP ☛

Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005 107

Nimetön 2 6 9/9/05 9:10:33 AM is more sensitive to variation in the visual rial. What that study showed, and what the Acknowledgments 17. Ren, G., Chen, F.: Determination of moisture con- region the spectra, when compared to the present work manifests, is that the bleach- tent of ginseng by near infra-red reflectance spectroscopy. Food Chemistry 60(1997):3, 433–436. peroxide bleached TMP. ing efficiency of dithionite is very depend- Grants from The Swedish Council for 18. Rantanen, J., Lehtola, S., Ramet, P., Mannermaa, ent on the condition of the outer part of the Forestry gratefully acknowledged. We are J.P., Yliruusi, J.: On-line monitoring of moisture PLS modelling using wavelengths sapwood, i.e. the wood closest to the bark. also grateful for the assistance provided content in an instrumented fluidized bed granu- between 400 to 1000 nm The efficiency of peroxide bleaching is less by the staff at Modo Paper Research & lator with a multi-channel NIR moisture sensor. Powder Technology 99(1998):2, 163–170. The wavelengths between 1002 and 2500 dependent on outer sapwood changes, such Development, and would like to thank Curt 19. Lorås, V.: Bleachability of mechanical pulp. Tappi nm were removed prior to PLS modelling as bark staining, and the bleaching efficiency Hägglund, Lars-Herman Sjöström and Bo Journal 57(1974):2, 98–102. in order to examine the model prediction is more influenced by the overall wood qual- Ahlqvist. 20. Elowsson, T., Liukko, K.: How to achieve effective ability when only using the wavelengths be- ity. These reported observations are only wet storage of pine logs (Pinus sylvestris) with a minimum amount of water. Forest Production tween 400 to 1000 nm. The reduced models valid for the actual bleaching procedures, References Journal 45(1995):11/12, 36–42. managed to model the ISO brightness of i.e. bleaching conditions closely reflecting 21. Sorge, C., Schulten, H.-R.: Influence of wet storage dithionite bleached TMP fairly well with those found in the pulp industry. 1. Wold, S., Johansson, E., Cocchi, M.: 3D QSAR of spruce wood on groundwater quality. Journal of RMSEP values (Table 1) somewhat higher The ISO brightness for peroxide Drug Design; Theory Methods and Applications. Environmental Analytical Chemistry 57(1994), Escom, Leiden 1993, p. 523. 1–8. than those achieved from the models using bleached as well as dithionite bleached TMP 2. Geladi, P., Kowalski, B.R.: Partial least squares 22. Clausen, C.A.: International Biodeterioration & the whole spectral region. The modelling was well correlated to all types of spectra regression – A tutorial. Analytica Chimica Acta Biodegradation 1996 p. 101. of the ISO brightness measured on perox- except for the modelling between the ISO 185(1986), 1–17. 23. Brown, D.G. and Abbot, J.: Effects of metal-ions ide bleached TMP did even better and gave brightness of peroxide bleached TMP and 3. Wold, S., Esbensen, K., Geladi P.: Principal Com- and stabilizers on peroxide decomposition during ponent Analysis. Chemometrics and Intelligent bleaching. Journal of Wood Chemistry and Tech- acceptable predictions for methods B, C spectra recorded on the sapwood closest to Laboratory Systems 2(1987), 37–52. nology 15(1995):1, 85–111. and D. The models computed from spec- the bark. The predictive ability that origi- 4. Geladi, P., MacDougall, D., Martens, H.: Lineariza- 24. Ju, Y., Ni, Y., Ohi, H.: Journal of Wood Chemistry tra measured on TMP produced a RMSEP nated from the near-infrared region was tion and scatter-correction for NIR reflectance and Technology 19(1999):4, 323–334. value of 0.74, which is lower than the cor- generally very low for the samples that had spectra of meat. Applied Spectroscopy 39(1985):3, 25. Abbot, J.: Catalytic decomposition of alkaline 491–500. hydrogen-peroxide in the presence of metal-ions responding value for the full models. a significant variation in water content. 5. Geladi, P., Martens, H.: Data pretreatment and – Binuclear complex formation. Journal of Pulp Although much of the irrelevant variation principal component regression using Matlab. and Paper Science 17(1991):1, 10–17. PLS modelling using wavelengths could be removed by using orthogonal sig- Journal of Near Infrared Spectroscopy 4(1996), 26. Leary, G., Giampaolo, D.: The darkening reactions between 1002 to 2500 nm nal correction, there was still variation left 225–242. of TMP and BTMP during alkaline peroxide 6. Wold, S., Antti, H., Lindgren, F., Öhman, J.: Or- bleaching. Journal of Pulp and Paper Science The wavelengths between 400 and 1000 in the spectra that was due to dissimilarities thogonal signal correction of near-infrared spectra. 25(1999):4, 141–147. nm were removed in order to find out if in water binding. It was however possible to Chemometrics and Intelligent Laboratory Systems 27. Bristow, J.A.: What is ISO brightness?. Tappi Jour- it is necessary to include the upper part of improve the correlation to the near-infrared 44(1998), 175–185. nal 77(1994):5, 174–178. the near-infrared region. For the dithionite region by drying the samples before per- 7. Trygg, J., Wold, S.: Orthogonal projections to latent 28. Bristow, J.A..: ISO brightness – A more complete structures, O-PLS. Journal of Chemometrics definition. Tappi Journal 82(1999):10, 54–56. bleached samples, the models based on the forming the spectroscopic measurements. 16(2002), 119–128. 29. Borga, P., Hämäläinen, M., Theander, O.: Cor- spectra measured according to Method C This lowered the prediction error signifi- 8. Fearn, T.: On orthogonal signal correction. Ch- relations between near-infrared spectra of wet- gave a RMSEP value of 1.12. This is very cantly for the models predicting ISO bright- emometrics and Intelligent Laboratory Systems stored timber and the storage time in relation similar to the corresponding value produced ness of dithionite bleached pulp from the 50(2000), 47–52. to the water-quality. Holzforschung 46(1992):4, 9. Antti, H., Sjöström, M., Wallbäcks, L.: Multivariate 299–303. from the full models and somewhat better near-infrared region of the spectra while the calibration models using NIR spectroscopy on 30. Nilsson, D., Edlund, U., Elg Kristofferson, K., than the value given by the models made predictions of ISO brightness for peroxide pulp and paper industrial applications. Journal of Sjöström, M., Agnemo, R.: The effect of designed on the wavelengths between 400 to 1000 bleached pulp based on the visual region Chemometrics 10(1996), 591–603. wood storage on the brightness of bleached and nm. The RMSEP value obtained from the of the spectra was almost kept at the same 10. Marklund, A., Paper, M., Hauksson, J.B., Sjöström, unbleached thermo mechanical pulp. Nordic M.: Prediction of strength parameters for softwood Pulp and Paper Research Journal 18(2003):4, models predicting peroxide bleached TMP level as before. kraft pulps. Multivariate data analysis based on 369–376. brightness was 0.85, which is slightly higher The three different kinds of data pre- orthogonal signal correction and near infrared 31. Wold, S.: Cross-validatory estimation of the number than for the previous models. Any modelling treatment techniques used in this study spectroscopy. Nordic Pulp and Paper Research of components in factor and principal components based on the other sample pre-treatment produced very similar results with respect to Journal. 14(1999):2, 140–148. models. Technometrics 20(1978), 397–404. 11. Malkavaara, P., Alén, R.: A spectroscopic method 32. Mathworks Inc. (www.mathworks.com), Natick, methods gave very poor results. prediction error. In some cases, the models for determining lignin content of softwood and United States 1997. using OSC tended to give slightly higher hardwood kraft pulps. Chemometrics and Intel- 33. Camo ASA (www.camo.com), Oslo, Norway ligent Laboratory Systems 44(1998), 287–292. Conclusions RMSEP values compared to the other two 1999. data pre-treatment techniques. This is re- 12. Wallbäcks, L., Edlund, U., Nordén, B., Berglund, I.: 34. Umetrics AB (www.umetrics.com), Umeå, Sweden Multivariate characterization of pulp using solid- 1999. The present modelling study showed that it flected in the RMSEP values for the models state C13 NMR, FTIR, and NIR. Tappi Journal 35. Spotfire AB (www.spotfire.com), Göteborg, Sweden is possible to build satisfactory PLS models based on the OSC pre-treated spectra of 74(1991):10, 201–206. 1999. between the ISO brightness of dithionite TMP, where model over fitting results in 13. Hiukka, R.: A multivariate approach to the analysis bleached TMP and the spectra measured lower prediction ability. of pine needle samples using NIR. Chemometrics and Intelligent Laboratory Systems 44(1998), directly on the wood surface. However, The final conclusion that can be made 395–401. Received January 30, 2004 for peroxide bleached TMP this was not from the PLS modelling is that the most 14. Durán, N., Angelo, R.: Infrared microspectroscopy the case. This does illustrate how the two important part of the spectra lies in the re- in the pulp and paper making industry. Applied bleaching methods compare with regards gion between 400 to 1000 nm. Prediction Spectroscopy Reviews 33(1998):3, 219–236. 15. Osborne, B.G., Fearn, T., Hindle, P.H.: Practical Appendix 1. to wood factors affecting the bleaching ef- models can in most cases be calculated solely NIR Spectroscopy with Applications in Food and ficiency. In a previous study /30/ 30, the fac- from this region. However, the inclusion of Beverage Analysis, Longman Group UK Limited, Bleaching tors controlling wood storage were examined the wavelengths between 1000 to 2500 nm Harlow 1993. 300 g (dry weight) from each pulp sample and it was found that the major difference does not worsen the predictive ability but 16. Derksen, M., Van de Oetelaar, P., Maris, F.: Ef- was treated with EDTA. The chelation re- ficient prediction of the residual moisture content between the efficiency of dithionite and per- can in fact in most cases improve the mod- specification of a freeze-dried product by using action was accomplished using plastic con- oxide bleaching was due to the interaction elling slightly, which is especially true for near-infrared reflectance spectroscopy. Journal of tainers heated with a water bath holding between light and irrigation, i.e. a factor the prediction of dithionite bleached TMP Pharmaceutical and Biomedical Analysis 17(1998), 70 °C. An excess of EDTA was used and mainly affecting the outer sapwood mate- brightness. 473–480.

108 Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005

Nimetön 2 7 9/9/05 9:10:35 AM Acknowledgments 17. Ren, G., Chen, F.: Determination of moisture con- pH was buffered to 5.5 during the reaction. Appendix 2. tent of ginseng by near infra-red reflectance spec- All pulp samples were washed and then troscopy. Food Chemistry 60(1997):3, 433–436. Grants from The Swedish Council for 18. Rantanen, J., Lehtola, S., Ramet, P., Mannermaa, bleached with hydrogen peroxide and sodi- Method A. Measurements directly on Forestry gratefully acknowledged. We are J.P., Yliruusi, J.: On-line monitoring of moisture um dithionite and each sample was bleached wood surface also grateful for the assistance provided content in an instrumented fluidized bed granu- twice using only single-stage bleaching with These measurements were carried out us- by the staff at Modo Paper Research & lator with a multi-channel NIR moisture sensor. either of the two methods at the same time. ing a fibre optic reflectance probe that was Powder Technology 99(1998):2, 163–170. Development, and would like to thank Curt 19. Lorås, V.: Bleachability of mechanical pulp. Tappi Thus, the bleached material consisted of 45 held tightly against the debarked peripheral Hägglund, Lars-Herman Sjöström and Bo Journal 57(1974):2, 98–102. peroxide bleached samples and 45 dithionite surface of the measured disc. Each disc was Ahlqvist. 20. Elowsson, T., Liukko, K.: How to achieve effective bleached samples. measured at three different spots, which re- wet storage of pine logs (Pinus sylvestris) with a sulted in that 12 measurements were done minimum amount of water. Forest Production Conditions for bleaching with H O for every sample (each sample consisted of References Journal 45(1995):11/12, 36–42. 2 2 21. Sorge, C., Schulten, H.-R.: Influence of wet storage Bleached pulp quantity: 50g (dry weight) four discs). The wavelength regions between 1. Wold, S., Johansson, E., Cocchi, M.: 3D QSAR of spruce wood on groundwater quality. Journal of Temperature: 70°C 2300 to 2500 nm were discarded due to Drug Design; Theory Methods and Applications. Environmental Analytical Chemistry 57(1994), Pulp concentration: 20% noise caused by the fibre optic instrumen- Escom, Leiden 1993, p. 523. 1–8. 2. Geladi, P., Kowalski, B.R.: Partial least squares 22. Clausen, C.A.: International Biodeterioration & Reaction time: 2h tation. The 12 spectra measured for each regression – A tutorial. Analytica Chimica Acta Biodegradation 1996 p. 101. Bleaching chemicals & conditions: H2O2 sample were checked for outliers (see data 185(1986), 1–17. 23. Brown, D.G. and Abbot, J.: Effects of metal-ions 4%, NaOH 3.7%, silicate 2.6%, intial pH treatment) and were then averaged to one 3. Wold, S., Esbensen, K., Geladi P.: Principal Com- and stabilizers on peroxide decomposition during ~11, end pH ~9 spectrum. ponent Analysis. Chemometrics and Intelligent bleaching. Journal of Wood Chemistry and Tech- Laboratory Systems 2(1987), 37–52. nology 15(1995):1, 85–111. The pulp was mixed with the bleaching 4. Geladi, P., MacDougall, D., Martens, H.: Lineariza- 24. Ju, Y., Ni, Y., Ohi, H.: Journal of Wood Chemistry chemicals in a mixer. The bleaching was car- Method B. Measurements on ground tion and scatter-correction for NIR reflectance and Technology 19(1999):4, 323–334. ried out in waterproof plastic bags heated wood spectra of meat. Applied Spectroscopy 39(1985):3, 25. Abbot, J.: Catalytic decomposition of alkaline in a water bath holding 70 °C. After two One fourth of each of the four debarked 491–500. hydrogen-peroxide in the presence of metal-ions 5. Geladi, P., Martens, H.: Data pretreatment and – Binuclear complex formation. Journal of Pulp hours of bleaching the pulp was washed in discs in each sample was chopped into small- principal component regression using Matlab. and Paper Science 17(1991):1, 10–17. a centrifuge with 6 litres of water and buff- er pieces. All knots and any bark material Journal of Near Infrared Spectroscopy 4(1996), 26. Leary, G., Giampaolo, D.: The darkening reactions ered to pH 5.5. were carefully avoided and excluded in the 225–242. of TMP and BTMP during alkaline peroxide sample material. The wood pieces were then 6. Wold, S., Antti, H., Lindgren, F., Öhman, J.: Or- bleaching. Journal of Pulp and Paper Science thogonal signal correction of near-infrared spectra. 25(1999):4, 141–147. Conditions for bleaching with Na2S2O4 ground in a Retch 2000 heavy-duty knife Chemometrics and Intelligent Laboratory Systems 27. Bristow, J.A.: What is ISO brightness?. Tappi Jour- Bleached pulp quantity: 35g (dry weight) mill with a rotational speed of 750 rpm. 44(1998), 175–185. nal 77(1994):5, 174–178. Temperature: 70% The ground wood was collected through a 7. Trygg, J., Wold, S.: Orthogonal projections to latent 28. Bristow, J.A..: ISO brightness – A more complete Pulp concentration: 4% 1.5 mm sieve and put in plastic bags. The structures, O-PLS. Journal of Chemometrics definition. Tappi Journal 82(1999):10, 54–56. 16(2002), 119–128. 29. Borga, P., Hämäläinen, M., Theander, O.: Cor- Reaction time: 30 min mill was thoroughly cleaned between every 8. Fearn, T.: On orthogonal signal correction. Ch- relations between near-infrared spectra of wet- Bleaching chemicals & conditions: Na2S2O4 grinding procedure. emometrics and Intelligent Laboratory Systems stored timber and the storage time in relation 2%, NaOH 0.2 M, initial pH ~11, end The NIR measurements were accom- 50(2000), 47–52. to the water-quality. Holzforschung 46(1992):4, pH ~7 plished using a spinning cup that was filled 9. Antti, H., Sjöström, M., Wallbäcks, L.: Multivariate 299–303. calibration models using NIR spectroscopy on 30. Nilsson, D., Edlund, U., Elg Kristofferson, K., A solution containing water, Na2S2O4 with ground wood and sealed with a card- pulp and paper industrial applications. Journal of Sjöström, M., Agnemo, R.: The effect of designed and NaOH was prepared prior to bleaching. board disc. All samples were measured three Chemometrics 10(1996), 591–603. wood storage on the brightness of bleached and A new solution was made for every third times and the cup was emptied and refilled 10. Marklund, A., Paper, M., Hauksson, J.B., Sjöström, unbleached thermo mechanical pulp. Nordic bleaching experiment to make sure that before a new measurement was started. The M.: Prediction of strength parameters for softwood Pulp and Paper Research Journal 18(2003):4, kraft pulps. Multivariate data analysis based on 369–376. the dithionite concentration was kept at a replicate spectra were checked for outliers orthogonal signal correction and near infrared 31. Wold, S.: Cross-validatory estimation of the number constant level. The pulp, containing 96% and were then averaged to mean spectra. spectroscopy. Nordic Pulp and Paper Research of components in factor and principal components water, was put in a glass beaker connected After that, the four spectra of each sample Journal. 14(1999):2, 140–148. models. Technometrics 20(1978), 397–404. to a nitrogen evacuation system, preventing were averaged to one spectrum. 11. Malkavaara, P., Alén, R.: A spectroscopic method 32. Mathworks Inc. (www.mathworks.com), Natick, for determining lignin content of softwood and United States 1997. degradation of dithionite by oxygen. The lid hardwood kraft pulps. Chemometrics and Intel- 33. Camo ASA (www.camo.com), Oslo, Norway of the container was equipped with an inlet Method C. Measurements on dried ligent Laboratory Systems 44(1998), 287–292. 1999. and outlet for nitrogen gas and an inlet for ground wood 12. Wallbäcks, L., Edlund, U., Nordén, B., Berglund, I.: 34. Umetrics AB (www.umetrics.com), Umeå, Sweden filling the beaker with dithionite solution. In order to achieve the same moisture con- Multivariate characterization of pulp using solid- 1999. state C13 NMR, FTIR, and NIR. Tappi Journal 35. Spotfire AB (www.spotfire.com), Göteborg, Sweden The beaker was evacuated three times prior tent for all samples the ground wood was 74(1991):10, 201–206. 1999. to bleaching in order to keep oxygen free dried in an oven for four hours at 105 °C. 13. Hiukka, R.: A multivariate approach to the analysis environment. Each evacuation started with The samples were then put in a dessicator of pine needle samples using NIR. Chemometrics filling the beaker with nitrogen gas followed and were allowed to cool to room tempera- and Intelligent Laboratory Systems 44(1998), 395–401. Received January 30, 2004 by evacuation of the beaker with a water ture. Apart from that, the dried samples were 14. Durán, N., Angelo, R.: Infrared microspectroscopy suction apparatus. The dithionite solution treated in the same way as for the milled in the pulp and paper making industry. Applied was then added to the beaker and a gentle samples described above. Spectroscopy Reviews 33(1998):3, 219–236. nitrogen gas flow was applied to the beak- 15. Osborne, B.G., Fearn, T., Hindle, P.H.: Practical Appendix 1. NIR Spectroscopy with Applications in Food and er. The temperature was kept constant at Method D. Measurements on thermo Beverage Analysis, Longman Group UK Limited, Bleaching 70 °C using a water bath covering the beak- mechanical pulp Harlow 1993. 300 g (dry weight) from each pulp sample er. The pulp was washed in a centrifuge with Measurements were also performed on the 16. Derksen, M., Van de Oetelaar, P., Maris, F.: Ef- was treated with EDTA. The chelation re- 6 litres of water after each bleaching experi- produced thermo mechanical pulp samples. ficient prediction of the residual moisture content specification of a freeze-dried product by using action was accomplished using plastic con- ment was complete. The solution was then The TMP measurements were accomplished near-infrared reflectance spectroscopy. Journal of tainers heated with a water bath holding buffered to pH 5.5. using the spinning cup as well. Pharmaceutical and Biomedical Analysis 17(1998), 70 °C. An excess of EDTA was used and 473–480.

Paperi ja Puu – Paper and Timber Vol.87/No. 2/2005 109

Nimetön 2 8 9/9/05 9:10:36 AM