UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Data mining in the spectro-microscopic analysis of complex material Tavassoli, Najmeh 2017

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


24-ubc_2017_september_tavassoli_najmeh.pdf [ 151.27MB ]
JSON: 24-1.0348732.json
JSON-LD: 24-1.0348732-ld.json
RDF/XML (Pretty): 24-1.0348732-rdf.xml
RDF/JSON: 24-1.0348732-rdf.json
Turtle: 24-1.0348732-turtle.txt
N-Triples: 24-1.0348732-rdf-ntriples.txt
Original Record: 24-1.0348732-source.json
Full Text

Full Text

Data Mining in the Spectro-Microscopic Analysis ofComplex MaterialbyNajmeh TavassoliM.Sc, University of Shahid Beheshti, 2010B.Sc, Sharif University of Technology, 2007A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Chemistry)The University of British Columbia(Vancouver)July 2017c© Najmeh Tavassoli, 2017AbstractVibrational spectroscopy has received significant interest in last decades as a ro-bust, rapid, and cost-effective alternative to the traditional wet-chemical methodsemployed by various industries. The spectra of complex materials may containsome components with a low concentration, whose information is buried within amajor peak of another component. These small hidden peaks contain critical infor-mation in some analysis. This thesis aims to develop novel data mining methods toimprove the quality of data, select its essential features, and finally build predictionmodels.The pulp industry offers one example in which spectroscopy offers attractiveadvantages as an on-line method for optimizing manufacturing. While spectro-scopic techniques are inherently sensitive to many of properties of interest to thepulp industry, they are potentially sensitive to provide features uncorrelated withphysical properties of pulp; which could hinder the development of robust pre-diction models. To overcome this challenge, we introduced Template Oriented Ge-netic Algorithm (TOGA). TOGA is aimed to establish significant features to assignpredictors according to a template determined to minimize prediction variance ina calibration space. It was found that TOGA significantly improved the predictionaccuracy of certain pulp properties compared to those without undergoing thesedata processing techniques.Near Infrared (NIR) is the most well-known spectroscopy technique which hasbeen successfully applied to pulp industry. However, broad overtone NIR absorp-tion band makes discerning of signature features a difficult process. We showedthat a combination of DWT and Orthogonal Signal Correction improved accuracyof prediction models built based on pulp NIR spectra.iiIn the second part of this thesis, a combined technique of interferometric scat-tering microscopy (iSCAT) and Raman spectroscopy was used to study the dynam-ics of gold-nanoparticle cellular uptake in cancer and normal cells model. Imagesderived from the study of these complex samples are heterogeneous which poses achallenge on true quantification and identification of the structure and componentsof a cell. To address this challenge, we used DWT to remove the out-of-focus anduncorrelated features from the original iSCAT images. This would make a true 3Dvolume of a cell and a precise track of AuNP internalizing a cell.iiiLay SummaryVibrational spectroscopy is a robust, rapid, and cost-effective alternative to thetraditional wet-chemical methods employed by various industries. Spectroscopictechniques show inherent sensitivity to many properties of interest, but do not yieldinformation with the same clarity as wet-chemical methods and thus must be pairedwith advanced data treatment techniques in order to yield useful results. This the-sis attempts to develop novel algorithms to improve the quality of data, select theessential features and build prediction models. In the first section of this thesis,a combination of vibrational spectroscopy and data processing technique is em-ployed to introduce a robust method to predict paper properties from on-line spec-tra of production pulps. In the second part, a combination of novel microscopytechnique known as iSCAT with modified image processing algorithm is used toprovide high resolution, real-time morphological and chemical information in threedimensions without the need for sample preparation.ivPrefaceThis thesis is made as a completion of the PhD education in Chemistry. Severalpersons have contributed academically and practically to this PhD thesis. First, Iwould like to thank my supervisor Dr. Edward Grant for his excellent guidance,valuable input and support throughout the entire my PhD period.Chapter 3 of this thesis has been published as N. Tavassoli, Z. Chen, A. Bain,L. Melo, D. Chen, and E. R. Grant, Template-Oriented Genetic Algorithm FeatureSelection of Analyte Wavelets in the Raman Spectrum of a Complex Mixture, An-alytical Chemistry, vol. 86, issue 21, 10591-10599. I was responsible for the datacollection, analysis as well as the manuscript composition. D. Chen was involvedin the chemometrics algorithm used in this study. E. R. Grant was the supervisoryauthor and was involved with concept formation and manuscript composition.Chapter 4 of this thesis has been published as N. Tavassoli and W. Tsai andP. Bicho and E. R. Grant, Multivariate classification of pulp NIR spectra for end-product properties using discrete wavelet transform with orthogonal signal cor-rection, Analytical Methods, vol. 6, issue 22, 8906–8914. I was responsible fordeveloping and modifying chemometrics algorithms, data analysis as well as themanuscript composition. W. Tsai and P. Bicho from Canfor industry were respon-sible in preparing pulp sheets and NIR data collection. E. R. Grant was the supervi-sory author and was involved with concept formation and manuscript composition.A part of chapter 5 of this thesis has been published in a google patent asMethod and apparatus for controlling a cellulosic pulp process, P.A. Bicho, E.R.Grant, P.W. Tsai, N. Tavassoli, Google patent, publication number WO2016090455A1, ”https://google.com/patents/WO2016090455A1?cl=tr”. I was responsible fordata collection, instrument modification, chemometrics and LabView algorithmsvdevelopement, and data analysis. W. Tsai and P. Bicho from Canfor industry as-sisted in Raman data collection, they were also responsible in preparing pulp sheetsand the manuscript composition. E. R. Grant was the supervisory author and wasinvolved with concept formation and manuscript composition.For chapter 6 which used the data in chapters 4 and 5, I was responsible forall major areas of concept formation, data analysis, as well as the majority ofmanuscript composition. Christy A. collected the ATR-IR data. A. Christy andA. Bain were contributed to manuscript edits. E. R. Grant was the supervisoryauthor and was involved with concept formation and manuscript composition.In chapter 7, I was responsible for all major areas of concept formation, datacollection, chemometrics algorithms modification, and data analysis, as well as themajority of manuscript composition. Ashton Christy and Qifeng Li were respon-sible in building instrument. Evan Shepherdson and Luke Melo assisted in datacollection. E. Shepherdson and A. Christy and were contributed to manuscript ed-its. E. Polishchuk in the biological services laboratory at UBC was responsible insample preparation. E. R. Grant was the supervisory author and was involved withconcept formation and manuscript composition.Works represented in appendix are in collaboration with different groups. Theresults shown in appendix 2 were in collaboration with Mahsa Imani under super-visory of Nando de Freitas from Computer Science department, UBC.viTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviList of Abbreviation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxviii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Outline of This Thesis . . . . . . . . . . . . . . . . . . . . . . . 42 Spectroscopy and Chemometrics . . . . . . . . . . . . . . . . . . . . 72.1 Vibrational Spectroscopy . . . . . . . . . . . . . . . . . . . . . . 72.1.1 The Main Principle . . . . . . . . . . . . . . . . . . . . . 72.1.2 Mid Infrared (Mid IR) . . . . . . . . . . . . . . . . . . . 102.1.3 Reflectance Methods ( Attenuated Total Reflectance Spec-troscopy) . . . . . . . . . . . . . . . . . . . . . . . . . . 102.1.4 Fourier Transform Infrared Spectrometer . . . . . . . . . 12vii2.1.5 Near Infrared (NIR) . . . . . . . . . . . . . . . . . . . . 132.1.6 Instrumentation . . . . . . . . . . . . . . . . . . . . . . . 142.1.7 Raman Spestroscopy and Its Instrumentation . . . . . . . 152.1.8 Resonance Raman Spectroscopy . . . . . . . . . . . . . . 162.1.9 Coherent Anti-Stokes Raman Spectroscopy (CARS) . . . 172.1.10 Stimulated Raman Scattering (SRS) . . . . . . . . . . . . 182.1.11 Surface Enhanced Raman Spectroscopy (SERS) . . . . . 192.1.12 Confocal Raman . . . . . . . . . . . . . . . . . . . . . . 202.2 Interferometric Scattering Microscopy (iSCAT) . . . . . . . . . . 232.2.1 Pure and Interferometric Scattering Microscopy . . . . . . 232.2.2 Disadvantages of Scattering Methodologies . . . . . . . . 252.2.3 Using Gold Nanoparticles to Address the Size Dependencyof Scattering Techniques . . . . . . . . . . . . . . . . . . 262.2.4 Addressing the High Dependency of Dark-Field on Parti-cle Size . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.2.5 History of Interferometric Scattering . . . . . . . . . . . . 272.2.6 Combined iSCAT and Confocal Raman Microscopy . . . 282.2.7 Limitations of iSCAT . . . . . . . . . . . . . . . . . . . . 312.3 Factors Affect the Quality of Images . . . . . . . . . . . . . . . . 312.3.1 Point Spread Function Mode . . . . . . . . . . . . . . . . 312.3.2 Aberration in Image Formation . . . . . . . . . . . . . . 322.3.3 Chromatic Aberration . . . . . . . . . . . . . . . . . . . 322.3.4 Spherical Aberration . . . . . . . . . . . . . . . . . . . . 332.3.5 COMA Aberration . . . . . . . . . . . . . . . . . . . . . 332.3.6 Astigmatism . . . . . . . . . . . . . . . . . . . . . . . . 342.3.7 Noise in Image Formation . . . . . . . . . . . . . . . . . 352.3.8 Shot Noise . . . . . . . . . . . . . . . . . . . . . . . . . 352.3.9 Thermal noise (Johnson-Nyquist Noise) . . . . . . . . . . 352.3.10 Dark Current Noise . . . . . . . . . . . . . . . . . . . . . 362.3.11 Readout Noise . . . . . . . . . . . . . . . . . . . . . . . 362.4 Wavelet Transfrom . . . . . . . . . . . . . . . . . . . . . . . . . 362.4.1 Fourier Transform (FT) . . . . . . . . . . . . . . . . . . . 362.4.2 Short-Time Fourier Transform (STFT) . . . . . . . . . . 38viii2.4.3 Continuous Wavelet Transform (CWT) . . . . . . . . . . 392.4.4 Discrete Wavelet Transform (DWT) . . . . . . . . . . . . 412.4.5 Two-Dimensional Wavelet Transform . . . . . . . . . . . 423 Template-Oriented Genetic Algorithm Feature Selection of AnalyteWavelets in the Raman Spectrum of a Complex Mixture . . . . . . . 443.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 473.2.1 Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.2.2 Spectroscopic Instrumentation and Measurement . . . . . 493.2.3 Spectral Analysis and Chemometric Analysis . . . . . . . 493.2.4 Spectrochemical Data Processing and Classification . . . 493.2.5 Discrete Wavelet Transformation . . . . . . . . . . . . . 493.2.6 Reconstructing a Feature-Selected Spectrum . . . . . . . 523.2.7 Multivariate Calibration . . . . . . . . . . . . . . . . . . 533.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . 543.3.1 Feature-Selected Spectra of Pure Solutions and RandomMixtures . . . . . . . . . . . . . . . . . . . . . . . . . . 543.3.2 Effect of Feature Selection on the Prediction Accuracy ofMultivariate Regression Models . . . . . . . . . . . . . . 573.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614 Multivariate classification of pulp NIR spectra for end-product prop-erties using discrete wavelet transform with orthogonal signal cor-rection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 644.2.1 Physical, Mechanical and Optical Properties of Pulp Sam-ples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644.2.2 Collection of NIR Spectra . . . . . . . . . . . . . . . . . 654.2.3 Multivariate Analysis . . . . . . . . . . . . . . . . . . . . 654.2.4 Data Pretreatment . . . . . . . . . . . . . . . . . . . . . . 654.2.5 First-Derivative Transformation . . . . . . . . . . . . . . 65ix4.2.6 Standard Normal Variant Correction . . . . . . . . . . . 664.2.7 Multiplicative Scatter Correction . . . . . . . . . . . . . 664.2.8 Discrete Wavelet Transform . . . . . . . . . . . . . . . . 674.2.9 Orthogonal Signal Correction . . . . . . . . . . . . . . . 674.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.3.1 NIR Spectra with Target Independent Pretreatment . . . . 704.3.2 Effects of Target-Directed Pretreatment . . . . . . . . . . 734.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734.4.1 Summary Overview of Pretreatment Effects . . . . . . . . 734.4.2 Prediction of Physical Properties on the Basis of NIR Clas-sification Models . . . . . . . . . . . . . . . . . . . . . . 744.4.3 Preprocessing as a Means of Isolating Spectroscopic Features 784.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825 TOGA Feature Selection and the Prediction of Physical-MechanicalProperties of Paper from the Raman Spectra of Unrefined Pulp . . . 845.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855.1.1 Chemical Composition of Pulp . . . . . . . . . . . . . . . 865.1.2 Pulp Morphology and Its Response to Refining . . . . . . 875.1.3 Prediction of Paper Properties from the Analysis of PulpComposition and Morphology . . . . . . . . . . . . . . . 875.2 Experimental methods . . . . . . . . . . . . . . . . . . . . . . . 895.2.1 Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . 895.2.2 Instrumentation . . . . . . . . . . . . . . . . . . . . . . . 905.2.3 Multivariate Analysis . . . . . . . . . . . . . . . . . . . . 915.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945.3.1 Physical and Mechanical Properties of Unrefined and Re-fined Pulp Sheets . . . . . . . . . . . . . . . . . . . . . . 945.3.2 Raman Spectra of Unrefined Pulp Sheets . . . . . . . . . 955.3.3 TOGA-Selected Sub-Spectra Associated with Particular Phys-ical Properties of Unrefined Pulp Sheets . . . . . . . . . . 965.3.4 TOGA-Selected Sub-Spectra Associated with Particular Phys-ical Properties of Refined Pulp Sheets . . . . . . . . . . . 99x5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1005.4.1 Pulp Properties as Affected by Fibre Structure and Fibre-Fibre Interactions . . . . . . . . . . . . . . . . . . . . . . 1005.4.2 Vibrational Manifestations of Pulp Fibre and Network Molec-ular Properties . . . . . . . . . . . . . . . . . . . . . . . 1025.4.3 TOGA-Selected Sub-Spectra Provide Evidence for Com-positional Factor Underlying Pulp Physical Properties ofUnrefined Pulp Sheet . . . . . . . . . . . . . . . . . . . . 1035.4.4 TOGA-Selected Sub-Spectra Provide Evidence for Com-positional Factors Underlying Pulp Physical Properties ofRefined Pulp Sheet . . . . . . . . . . . . . . . . . . . . . 1055.4.5 Multivariate Classification of Pulp Raman Spectra for End-Product Properties Using Template Oriented Genetic Al-gorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 1075.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1106 Combination of Multiple Spectroscopy Techniques Using Data Fu-sion for Enhanced Prediction Modeling of Physical-Mechanical Prop-erties of Paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1136.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1146.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 1186.2.1 Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . 1186.2.2 Spectroscopic Instrumentation and Measurement . . . . . 1186.2.3 Discrete Wavelet Transform . . . . . . . . . . . . . . . . 1196.2.4 Orthogonal Signal Correction . . . . . . . . . . . . . . . 1196.2.5 Template Oriented Genetic Algorithm . . . . . . . . . . . 1196.2.6 Data Fusion . . . . . . . . . . . . . . . . . . . . . . . . . 1206.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1236.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1266.4.1 Comparing the Efficiency of Spectroscopic Techniques inPredicting Physical Properties of Pulp . . . . . . . . . . . 1276.4.2 Fusing Spectroscopic Methods to Build a Final PredictionModel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133xi6.4.3 Fusing Spectroscopic Methods to Build the Final Predic-tion Model After Reducing the Number of Features . . . . 1346.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1397 Live, Three-Dimensional Dynamics of Nanoscale Particle Internal-ization, Detected Chemically and Morphologically in Human Cells . 1407.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1407.2 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . 1457.2.1 iSCAT-Raman Instrumentation . . . . . . . . . . . . . . . 1457.2.2 Sample Preparation . . . . . . . . . . . . . . . . . . . . . 1477.2.3 Image Processing . . . . . . . . . . . . . . . . . . . . . . 1487.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1547.3.1 Nanoparticle Dynamics Before, During, and After Inter-nalization by Live Cells . . . . . . . . . . . . . . . . . . 1547.3.2 Chemistry of Nanoparticle-Dense Areas Within Cells . . . 1577.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1587.4.1 Three-Dimensional Reconstruction of Cells . . . . . . . . 1597.4.2 Gold Nanoparticle Tracking . . . . . . . . . . . . . . . . 1627.4.3 Dynamics After Adding Gold Nanoparticles to Cell Samples1647.4.4 Raman Spectra . . . . . . . . . . . . . . . . . . . . . . . 1657.4.5 Comparison Between iSCAT-Raman and Other Techniques 1667.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1718 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176xiiList of TablesTable 2.1 Comparison of potentials, advantages and disadvantages of threedifferent vibrational spectroscopic techniques (Raman, MIR, andNIR). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22Table 3.1 Regression quality and prediction accuracy of multivariate mod-els for target sugar in mixtures with varying backgrounds ofother sugars based on complete and feature-selected DWT Ra-man spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . 58Table 3.2 Residual mean error of prediction observed in applying individ-ual TOGA models keyed to glucose, galactose, mannose andarabinose to five aqueous validation mixtures of these sugarscontaining the random concentrations listed below. . . . . . . 60Table 4.1 The experimental range of measurement and observed repro-ducibility of different the physical and optical properties inde-pendently determined for the samples that served as calibrationand validation standards for this study. . . . . . . . . . . . . . 64Table 4.2 Optimized number of PLS factors, h, and the regression coeffi-cient btained from the model fit to the calibration data-set, fol-lowing various methods of NIR spectral pretreatment. . . . . . 76xiiiTable 4.3 Average residual mean square errors of prediction following theapplication of various methods of NIR spectral data pretreat-ment. The RMSEP averages determined in each case by partialleast squares calibration models using ten different randomlyselected training and test data sets. Tabulated standard devia-tions reflect the reproducibility of RMSEP following differentpreprocessing methods. . . . . . . . . . . . . . . . . . . . . . 79Table 5.1 Prediction error for all physico-mechanical properties of pulpsamples using treated and untreated Raman spectra. . . . . . . 112Table 6.1 Comparison of the root mean square error of cross validation(RMSECV) of individual prediction models made with DWT-filtered data. Tensile, burst, tear, SRE, and density have beenlabeled in different freenesses as 600 ml, 500 ml, 450 ml, and300 ml. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130Table 6.2 Comparison of the root mean square error of cross validation(RMSECV) of prediction models made with individual spec-troscopic and data fusion techniques using preprocessed spec-tra with DWT. Tensile, burst, tear, SRE, and density have beenlabeled in different freenesses as 600 ml, 500 ml, 450 ml, and300 ml. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135Table 6.3 Comparison of prediction models made with individual spec-troscopic and data fusion techniques. In this table, Raman spec-tra have been treated with TOGA and NIR spectra with OSC,as chemometric feature selection methods. Tensile, burst, tear,SRE, and density have been labeled in different freenesses as600 ml, 500 ml, 450 ml, and 300 ml. . . . . . . . . . . . . . . 137Table A.1 The experimental range, average and the standard deviation ofmeasurment of different physical- mechanical properties of pulpsamples that served as calibration and validation standards inchapter 3-5. . . . . . . . . . . . . . . . . . . . . . . . . . . . 211xivTable B.1 Root mean square error all the regressors when they used theselected features and extracted features together. . . . . . . . . 213xvList of FiguresFigure 2.1 (left) harmonic model and (right) anharmonic model for thepotential energy of a diatomic molecule. . . . . . . . . . . . . 9Figure 2.2 Schematic of a multiple reflection in the crystal of high refrac-tive index in ATR instrument . . . . . . . . . . . . . . . . . . 11Figure 2.3 Schematic diagram of a Michelson interferometer used in FTIRtechnique. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12Figure 2.4 Three types of scattering by a molecule excited by a photonwith energy, E = hϑ . . . . . . . . . . . . . . . . . . . . . . . 15Figure 2.5 Raman spectra of fiber samples with large fluorescent back-ground. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16Figure 2.6 A simplified schematic diagram showing light absorption Stokesand anti-Stokes Resonance Raman scattering (RRS) processes. 17Figure 2.7 Energy level diagrams describing, (left) spontaneous Ramanscattering processes, (middle) stimulated Raman scattering (SRS),and (right) coherent anti-Stokes Raman scattering (CARS). . . 18Figure 2.8 Diagram of stimulated Raman scattering principle. It describesthat SRS signal is generated when a specific vibrational re-sponse based on the energy exchange between stokes and thepump light beams. . . . . . . . . . . . . . . . . . . . . . . . 19Figure 2.9 A schematic of conventional (left) and confocal (right) Ramaninstrument. In Confocal Raman just focused scattered lightwould be detected by detector. . . . . . . . . . . . . . . . . . 21xviFigure 2.10 Schematic diagram of the combined iSCAT-Raman microscope,including lasers operating at 532 nm (iSCAT) and 633 nm (Ra-man), acousto-optic beam deflectors (AOBD), long-pass filters(l p f ), CMOS cameras for iSCAT and Bright-field, confocalapparatus, and spectrograph. Note that the bright-field systemis peripheral and was not used in the present research. . . . . . 30Figure 2.11 Schematic diagram of chromatic aberration of a single lens . . 33Figure 2.12 Schematic diagram of spherical aberration of a single lens . . 33Figure 2.13 Schematic diagram of coma aberration of a single lens . . . . 34Figure 2.14 Schematic diagram of astigmatism effect . . . . . . . . . . . 34Figure 2.15 Illustration of the decomposition of an apparently random sig-nal (left) to its principal underlying frequencies at f1 = 30 Hzand f2 = 80 Hz (right). . . . . . . . . . . . . . . . . . . . . . 37Figure 2.16 Examples of mother function of wavelet suited for signal witha) abrupt discontinuities , b) smooth oscillations . . . . . . . . 40Figure 2.17 Multiresolution decomposition of a Raman signal of a pulpsample using symlet 5 as wavelet function . . . . . . . . . . . 42Figure 3.1 (left) α-D-pyranose structures of monosaccharides, glucose,galactose, mannose and arabinose. (right) Raw Raman spectraof pure monosaccharides in aqueous solution at concentrationsof 0.03 g ml−1, from the bottom, glucose, galactose, mannoseand arabinose. . . . . . . . . . . . . . . . . . . . . . . . . . 48xviiFigure 3.2 Mean-centred, normalized Raman spectra of a) 20 aqueous so-lutions of pure glucose over a range of concentrations from0.04 to 0.10 g ml−1. b) all 240 aqueous solutions of glucosegalactose, mannose and arabinose mixed in random amountsover a range of concentrations from 0.04 to 0.10 g ml−1. c)spectra reconstructed from the DWT-TOGA feature selectionof a). d) spectra reconstructed from the DWT-TOGA featureselection of the 60 sample calibration set for glucose with ran-dom additions of galactose, mannose and arabinose (Figure3.3b). Spectra, plotted by pixel number represent a Ramanshift interval from 320 to 2250 cm−1. . . . . . . . . . . . . . 55Figure 3.3 Mean-centred, normalized Raman spectra of a) 20 pure glu-cose solutions in a range of concentrations from 0.04 to 0.10g ml−1. b) 60 glucose solutions in a range of concentrationsfrom 0.04 to 0.10 g ml−1 and random added amounts of galac-tose, mannose and arabinose. c) and d) spectra reconstructedfrom DWT-TOGA feature selection of a) and b). . . . . . . . 56Figure 3.4 PLS model using a) raw Raman spectra, b) TOGA reconstructedspectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57Figure 3.5 Toga reconstructed spectrum for a) Glucose, b) Galactose, c)Mannose, d) Arabinose . . . . . . . . . . . . . . . . . . . . . 59Figure 4.1 (a) Representative NIR spectrum in cm−1 for each of the 75pulp samples in the present dataset (c.f. Table 4.1) (b) Spec-tral data subjected to first-derivative transformation. Spectraldata subjected to nnnnnnn(c) standard normal variant (SNV)transformation and (d) multiplicative scatter correction (MSC). 71xviiiFigure 4.2 (a) Spectra following a discrete wavelet transform of the rawspectral data in Figure 4.1 (b) The same spectra following or-thogonal signal correction (OSC), targeted to a prediction ofthe paper sheet mechanical property, tensile strength. (c) Rawspectral data pretreated first by discrete wavelet transform, asabove, followed by two-component orthogonal signal correc-tion targeted to tensile strength. (d) Pretreated spectral datareversing the order of transformation by OSC and DWT. . . . 72Figure 4.3 Pulp NIR spectra after processing by combination of orthogo-nal signal correction with two components and discrete wavelettransform with selected wavelet range of (3-7) with referenceto (left) Burst, and (right) Tensile Strength. . . . . . . . . . . 80Figure 4.4 Pulp NIR spectra after processing by combination of orthogo-nal signal correction with two components and discrete wavelettransform with selected wavelet range of (3-7) with referenceto (left) Wet Zero Span, and (right) Dry zero span. . . . . . . 81Figure 4.5 Pulp NIR spectra after processing by combination of orthogo-nal signal correction with two components and discrete wavelettransform with selected wavelet range of (3-7) with referenceto (left) SRE-600, and (right) Tear. . . . . . . . . . . . . . . . 81Figure 4.6 Pulp NIR spectra after processing by combination of orthogo-nal signal correction with two components and discrete wavelettransform with selected wavelet range of (3-7) with referenceto (left) Scattering, and (right) Freeness. . . . . . . . . . . . . 82Figure 5.1 Histogram showing tensile and tear properties of six differentof unrefined and refined pulp sheets, freeness= 600, 550, 500,450, 300. a) tensile, b) tear. . . . . . . . . . . . . . . . . . . . 95Figure 5.2 The correlation between tensile strength of six different of un-refined and refined pulp sheets, freeness= 600, 550, 500, 450,300. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96xixFigure 5.3 Pulp Raman spectra before/after preprocessing by discrete wavelettransform with selected wavelet range of (2-6) a) Raw Ramanspectra, b)DWT Raman spectra. . . . . . . . . . . . . . . . . 97Figure 5.4 Pulp Raman spectra after discrete wavelet transform and TOGAfeature selection guided to optimize the following properties ofunrefined pulp a) tensile strength, b) SRE600, c) burst strength,d) dry zero span, e)tear strength, f) wet zero span. . . . . . . . 98Figure 5.5 Raman spectra of unrefined pulps after wavelet transform andTOGA preprocessing to optimize the prediction of burst strength,tensile strength, tear strength and SRE600 at freeness of (a)600 ml, (b) 450 ml and (c) 300 ml . . . . . . . . . . . . . . . 99Figure 5.6 PLS models for untreated and treated Raman spectra by TOGAwith latent variable equals to 7 a)Tensile (untreated Raman) b)Tensile (treated Raman) c) Burst (untreated Raman) d) Burst(treated Raman). . . . . . . . . . . . . . . . . . . . . . . . . 108Figure 5.7 PLS models for untreated and treated Raman spectra by TOGAwith latent variable equals to 7 a)Tear (untreated Raman) b)Tear (treated Raman) c) SRE (untreated Raman) d) SRE (treatedRaman). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109Figure 6.1 A schematic chart of different levels of data fusion, a)Lowlevel, b)Medium level, and c)High level. . . . . . . . . . . . . 121Figure 6.2 A schematic chart of high level data fusion to calculate thefinal predicted property of interest. As the chart represents, wecalculate W as the weight of each technique to then fuse theresult of each techniques as the final result. . . . . . . . . . . 122Figure 6.3 Representative raw ATR, Raman, and NIR spectra in cm−1 foreach pulp samples in the present datasets. a) Raw ATR-FTIRspectra. b) Raw Raman spectra d) Raw NIR spectra. . . . . . 124Figure 6.4 Representative DWT treated ATR, Raman, and NIR spectra forpulp samples. a) DWT-treated ATR-FTIR spectra. b) DWT-treated Raman spectra c) DWT-treated NIR spectra . . . . . . 125xxFigure 6.5 Concatenation of DWT-processed NIR, ATR-FTIR, and Ra-man data sets for low-level fusion. . . . . . . . . . . . . . . . 126Figure 6.6 The calculated weight for each spectroscopic technique, whenconstructing a high-level data fusion model for different prop-erties. Red: Raman, Green: NIR, and Blue: ATR-FTIR . . . . 128Figure 6.7 Concatenation of DWT-OSC processed NIR, and TOGA pro-cessed Raman data sets for mid-level fusion. . . . . . . . . . 136Figure 6.8 The calculated weight for OSC-DWT-NIR and TOGA-Ramanspectra in building a high-level data fusion model for differentproperties. Tensile, burst, tear, SRE, and density have beenlabeled in different freenesses as 600 ml, 500 ml, 450 ml, and300 ml. Red: Raman, and Green: NIR. . . . . . . . . . . . . 138Figure 7.1 Schematic diagram of the combined iSCAT-Raman microscope,including lasers operating at 532 nm (iSCAT) and 633 nm (Ra-man), acousto-optic beam deflectors (AOBD), long-pass filters(l p f ), CMOS cameras for iSCAT and Bright-field, confocalapparatus, and spectrograph. Note that the bright-field systemis peripheral and was not used in the present research. . . . . 147Figure 7.2 (a): Gaussian wavelet used for 2-D Discrete Wavelet Trans-form (DWT). (b): iSCAT image of Hela and fibroblast cells, af-ter applying 2-D DWT to discrete image in different frequencylevels. The first-level low-low subimage (LL1) has been pro-cessed with 2-D DWT again to produce second-level subimages.150Figure 7.3 Graphical representations of edge structures. (a): Dirac struc-ture (b): Roof structure (c): A-step structure (d): G-stepstructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151Figure 7.4 Successive frames showing the interaction of two gold nanopar-ticles with the membrane of HeLa cell before internalization.Only a small part of the cell membrane is shown. The temporalinterval of each frame was 10 s. The scalebar is 200 nm. . . . 155xxiFigure 7.5 Gold nanoparticle dynamics in HeLa cell membranes. Top:Overall image of nanoparticles and cells. Bottom: Framesshowing the motion of nanoparticles inside the cell membrane;expansion of the illustrated area in the top image. The tempo-ral interval of each frame is 7 s.The scalebar for top and bottomimages is 10 µm and 200 nm, respectively. . . . . . . . . . . 156Figure 7.6 Surface enhanced Raman Spectra of different parts of a HeLacell. Exposure time: 0.5 s. Laser wavelength: 633 nm. black:lipid, blue: nucleic acid, red: protein, and green: carbohydrate. 158Figure 7.7 iSCAT image of a fibroblast cell. Left: Raw iSCAT image.Right: Deconvolved iSCAT image using DWT. The scale baris 10 µm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159Figure 7.8 HeLa cells affixed to the cover slip, after removing backgroundand noise. This image has been obtained by calculating thesum of 15 different z planes (0− 10 µm), showing HeLa cellstructure in the x, y, and z dimensions. . . . . . . . . . . . . . 160Figure 7.9 Processed iSCAT images of detached Hela cells. Left: Three-dimensional structure of five detached HeLa cells, with theirnuclei in red. Right: An xz section of the left-hand image, withnuclei in red, and the rest sections of cell in blue. . . . . . . . 161Figure 7.10 Three-dimensional gold nanoparticle trajectory while interact-ing with the exterior of a HeLa cell membrane (100 points,33 ms per point), reconstructed from the time-resolved iSCATimages with a temporal resolution of 33 ms. . . . . . . . . . . 163Figure 7.11 Three-dimensional particle trajectory while (left) passing throughthe membrane and (right) moving within the HeLa cell (1,200points, 33 ms per point), both reconstructed from the time-resolved iSCAT images with a temporal resolution of 33 msec.The scalebar is 200nm. . . . . . . . . . . . . . . . . . . . . . 163Figure 7.12 HeLa cell after culturing with 30 nm gold nanoparticles for 4hours. This image is a sum of 12 different z sections. . . . . . 164Figure 7.13 Four different z sections of HeLa cells after culturing with 30nm GNP for 4 hours. the depth is 8 µm with ∆z as 2 µm . . . 165xxiiFigure 7.14 Interactions of gold nanoparticles with HeLa (lower left) andfibroblast (right) cells. Left: Map of first principal componentof Raman data; orange areas are the most highly correlatedwith gold nanoparticles. Right: iSCAT image showing cells,before gold nanoparticles were added. . . . . . . . . . . . . . 166Figure A.1 The correlation between different mechanical properties of pulp-sheets made of a slightly beaten pulp with freeness 600 . . . . 210Figure C.1 A deconvolved iSCAT image of a single fibre in a pulp sample 214Figure C.2 DWT deconvolved iSCAT images of a single fiber in differentstacks containing information of 24 micron in Z direction, 20micron in X and 60 micron in Y direction. . . . . . . . . . . . 215Figure C.3 A reconstructed volume using the deconvolved stacks in Figurefig:fibre. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216Figure C.4 DWT deconvolved iSCAT images of a single fiber in differentstacks containing information of 20 micron in Z direction, 20micron in X and 60 micron in Y direction. . . . . . . . . . . . 217Figure C.5 A reconstructed volume using the deconvolved stacks in Figurefig:fibre2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218Figure C.6 DWT deconvolved iSCAT images of a pulp sample represent-ing the network of fibres in different stacks containing infor-mation of 0 micron in Z direction, 60 micron in X and 60 mi-cron in Y direction. . . . . . . . . . . . . . . . . . . . . . . . 219Figure D.1 3D iSCAT image of an olivin sample . . . . . . . . . . . . . 220Figure D.2 Raman spectra of an olivine sample contains fluid. Green spec-trum shows the Raman spectrum of olivine while red and bluespectra represent area of the sample with a hole. We were in-terested to find the composition of the fluid which fills up thehole. The results show that it is possible that the hole is emptyor the liquid has a low Raman cross section which hasn’t beenshown up in Raman spectrum. . . . . . . . . . . . . . . . . . 221xxiiiFigure D.3 left figure shows iSCAT image of an olivine sample with olivinein red and holes in blue and yellowish colors. right figure rep-resents Raman map of around the same area of the olivine sam-ple which we used to take iSCAT image. PC1 was used to buildthe present Raman map. . . . . . . . . . . . . . . . . . . . . 222Figure D.4 Left shows the iSCAT image of surface of an olivine sample.Right figure represent the same area of olivine sample in astack with 5 micron deeper. It shows that the holes in surfaceget smaller in deeper regions . . . . . . . . . . . . . . . . . . 223Figure E.1 This image shows the surface of a sample of mouse’s spinalcord representing a neuron. . . . . . . . . . . . . . . . . . . . 224Figure F.1 iSCAT images of different parts of an elastin sample. . . . . . 225Figure F.2 iSCAT images of an elastin sample with the calcificied crystals. 226Figure F.3 iSCAT images of an elastin sample with the calcificied crys-tals, the calcificied crystals are shown up in brighter color com-pared to elastin. . . . . . . . . . . . . . . . . . . . . . . . . 227Figure F.4 Top figure shows the iSCAT image of an elastin sample withcalcificied crystals, the bottom image serves Raman map of thesame sample around the same region which iSCAT image hasbeen taken. PC1 was used to make Raman map. . . . . . . . . 228Figure G.1 This image shows PDA-primed silicone surfaces. The thick-ness of PDA coating and the surface morphology are easilycan be detected . . . . . . . . . . . . . . . . . . . . . . . . . 230Figure G.2 This image shows PDA-primed silicone surfaces loaded withgentamicin. Comparing between this figure and above figureshows the different surface morphology before and after load-ing drug. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231Figure H.1 A DWT deconvolved iSCAT image of heLa cell using the stan-dard deviation of 25 Z stacks. . . . . . . . . . . . . . . . . . 232Figure H.2 DWT deconvolved iSCAT images of a heLa cell sample in fif-teen different depth. . . . . . . . . . . . . . . . . . . . . . . 233xxivList of Abbreviation1D One Dimensional2D Two DimensionalAOTF Acoustic Optic Tunable FilterAOD Acousto Optic beam DeflectorATR Attenuated Total ReflectionBVE Backward Variable EliminationCSM Canadian Standard MethodCARS Coherent Anti-Stokes Raman SpectroscopyCTW Continuous Wavelet TransformDIC Differential Interference ContrastDPSS Diode-Pumped Solid StateDWT Discrete Wavelet TransformDZS Dry Zero SpanDMEM Dulbeccos Modied Eagles MediumEPR Enhanced Permeability and RetentionFPALM Fluorescence Photoactivation MicroscopyFT Fourier TransformFT-IR Fourier Transform InfraredGA Genetic AlgorithmAuNPs Gold NanoparticlesISE Interactive Stepwise EliminationIRM Interference Reflection MicroscopyiSCAT interferometric Scattering microscopy (iSCAT)IPW Iterative Predictor WeightingxxvIVS Iterative Variable SelectionKDD Knowledge Discovery in DatabasesLSFM Laser Scanning Fluorescence MicroscopyLSM Light Sheet MicroscopyMIR Mid InfraredMCCV Monte Carlo cross validationMPEF Multiple-Photon Excitation Fluorescence MicroscopyMSC Multiplicative Scatter CorrectionNIR Near InfraredNBSK Northern Bleached Softwood KraftOSC Orthogonal Signal CorrectionPLS Partial Least SquaresPALM Photoactivated Localization MicroscopyPCA Principal Component AnalysisRRS Resonance Raman scatteringRMSECV Root Mean Square Error of Cross ValidationRMSEP Root Mean Square Error of PredictionSEM Scanning Electron MicroscopySTFT Short-Time Fourier TransformSNR Signal to Noise RatioSRE Specific Refining EnergySRG Stimulated Raman GainSRL Stimulated Raman LossSNV Standard Normal VariantSVS Stepwise Variable SelectionSTED Stimulated Emission Depletion MicroscopySRS Stimulated Raman ScatteringSTORM Stochastic Optical Reconstruction MicroscopySIM Structured Illumination MicroscopySERS Surface Enhanced Raman SpectroscopyTOGA Template Oriented Genetic AlgorithmTEM Transmission Electron MicroscopyxxviTNF Tumor Necrosis FactorUVE Uninformative Variable EliminationVCO Voltage Controlled OscillatorsWZS Wet Zero SpanxxviiAcknowledgmentsI would like to thank all the people who committed in different ways to my Ph.D.research which has been represented in this thesis. First and foremost, I thank mygraduate supervisor, Professor Edward Grant, for going beyond the role of a su-pervisor and being a friend and a mentor. In past years, he gave me intellectualfreedom in my research and in following my passion, supported my attendanceat various conferences, engaged me in new ideas, and provided me outstandingguidance and exciting projects. Additionally, Next, I would like to thank my com-mittee members Professor Roman Krems, Professor David Chen, and ProfessorAllan Bertram for their supports and helps in my Ph.D. research. My thanks alsogoes to my fellow lab- mates and collaborators, Dr. D. Chen, Dr. Zh. Chen, A.Christy, E. Shepherdson, L. Melo, Dr. H. Sadeghi, A. Bain, M. Schulz-Weilingand M. Aghigh. Their help and supports have big effects on the described resultsin this thesis were accomplished. Dr. Daniel Chen and I worked together on sev-eral different chemometric projects, and without his keen scientific insights, myjob would be more difficult. Last but not least, I want to thank my parents, my sib-lings, and my husband, whose everlasting support and love has always motivatedme to pursue my passion.xxviiiChapter 1IntroductionWe are drowning in information but starved for knowledge.John NaisbittIn an analytical process, developing and expanding of various instrumentscause a dramatic increase in analytical information, which highlights the neces-sity of a tool designed to explore a large amount of data [202, 324]. One of the fastdeveloping analytical techniques is vibrational spectroscopy, which is a fast non-destructive technique with no need for sample preparation [27, 47]. Vibrationalspectroscopy is a premier method of performing both qualitative and quantitativeanalysis in different area such as food [122, 134, 174, 185, 238, 254, 260, 283, 337],pharmacy [108, 155, 243, 259], biomedical [169, 233, 234, 258], and pulp and pa-per [4, 34, 74, 74, 89, 112, 113, 193, 210, 211] industry. However, spectroscopicdata, known as spectrum, may pose difficulties in data analysis as they consist morethan thousands spectral variables (frequencies) which could be potentially corre-lated. Each of these variables may refer to some chemical, physical, and structuralinformation of distinct components in a sample, which makes an interpretationof the spectrum a complex act. Even if it could be easily doable to analyze aspectrum of component, it could be difficult to distinguish between two spectraof components with similar chemical structures, when the spectra may have slightdifferences [21, 35, 108, 117, 146, 169]. To reduce the complexity of data analysis,data mining known as knowledge discovery in databases (KDD) is introduced tochemistry in 1998 by F.K. Brown [41]. This mathematical technique offers a fast,1cheap and efficient promise in exploring patterns between variables, getting hiddeninformation, and validating prediction models. This thesis aims at developing andintroducing some novel data mining methods which cause significant improvementin quantitative and qualitative analysis of spectro/microscopic measurements.Data mining employs three consequent steps to accurately treat and analyzespectroscopic data. These steps are (i) signal/image processing, (ii) feature selec-tion, and (iii) classification models. In the signal processing step, one generallyuses an appropriate preprocessing method to remove noise, reduce uncorrelatedvariances, and correct interferences such as baseline drift, path length variation inspectroscopic data. Various signal processing has been introduced in analyticalchemistry in last decades including first and second derivative [7, 12, 102, 115],Standard Normal Variant Correction (SNV) [29], Multiplicative Scatter Correc-tion (MSC) [13, 100, 194, 195], Orthogonal Signal Correction (OSC) [301, 347],Fourier Transform [9, 12, 48], and Discrete Wavelet Transform (DWT) [49, 80,148, 189, 307, 329]. DWT is one of the most efficient signal processing method,as it provides information in both frequency and time domain. Similar to FourierTransform, the DWT decompose a signal to different frequencies, while keepinginformation about the temporal location of the frequencies. So, each feature inspectrum can be studied with any resolution of interest. Consequently, few highestfrequencies (spectral noise) and a lowest frequency (background) can be removedfrom a raw spectrum. This leaves the midrange frequencies, the features of in-terest in the spectrum, undisturbed [49, 80, 329]. The second step is a properfeature selection technique; an efficient technique capable of exploring essentialfeatures representing the input pattern and decreasing the data dimension by ex-tracting the most important features. Different feature selection algorithms havebeen applied in different fields of analytical chemistry to explore and extract a setof features which has the most correlation with a property of interest. The selectionmechanism is based on excluding useless features according to their coefficients ina regression. Standard techniques, which represent model-wise elimination ap-proaches [93], include the iterative variable selection (IVS )[177], iterative predic-tor weighting (IPW)[50, 183], uninformative variable elimination (UVE)[43, 46],GOLPE [30, 60], and interactive stepwise elimination (ISE)[92, 93], stepwise vari-able selection (SVS) [56], or backward variable elimination (BVE) [180, 304],2and stepwise interval partial least-squares (backward or forward:BiPLS or FiPLS)[200, 225] . These methodologies make strong assumptions about the spectral pa-rameters, such as peak widths, and this can give rise to problems when the characterof the spectra is unknown. Genetic algorithm, which applies the principle of nat-ural selection to choose features without making any assumption about the searchspace [171, 197, 220], provides a tool to resolve this limitation.The third step is building a robust representative and reliable calibration/pre-diction model based on the important correlated features, known as classificationmodels. These models use an appropriate multivariate algorithm to link spectra toproperties of sample or concentration of compounds. Conventional spectroscopicdeterminations construct linear regression models based on univariate analysis, i.e.the relative intensity of a single feature [2, 332]. However, substances in a mixtureoften are presented by signals at different Raman scattered wavelengths. Multivari-ate analysis of the spectrum yield far better results for mixtures of varying compo-sition. Thus, multivariate statistical methods are suggested as a decisive techniqueto link the spectrum to quantifiable properties of samples [2, 62, 236, 323]. In somecases, the lack of information comes from one source makes multivariate analysisfail to construct an accurate prediction/calibration model. Interestingly, the combi-nation of several techniques has been successful to provide adequate informationin order to make a classification model. This process, commonly known as datafusion, is the integrating of multiple data sets representing the same real-worldproperty into an accurate representation is presented in several published works[37, 75, 106, 201, 339, 341]. In the second part of this thesis, we implementeda relatively novel technique known as interferometric scattering microscopy (iS-CAT) to get wide field images of biological cells. Here, we apply similar datamining steps to improve the resolution and contrast of an image (two-dimensionaldata). Fast and digital wide field iSCAT microscopy allows constructing a 3-D vol-ume of a cell out of many 2D stacks [57, 208, 229]. However, two factors limitthe resolution and reliability of making the cell volume: noise, and unwanted fea-tures which may be related to out of focus components and point spread function[140, 198, 229, 291].The first step to address the limitation of microscopy is preprocessing of rawimages. Images obtained by iSCAT mostly suffer from dominant background fea-3tures and high frequency shot noises. To improve the contrast and quality of theimage, the DWT approach is a suitable technique, as stated in 1D signal process-ing [12, 48, 72, 81]. The second step is to detect and remove out-of-focus fea-tures. Each image obtained by wide-field microscopy contains both sharp in-focusfeatures and blur features originated from the neighbor planes. Previous studiesshowed that blurriness mainly affects the type and sharpness of edges in an image[179]. Hence in this study, we utilized a method based on DWT to detect blurryand sharp edges. After detecting the blurred edges, they can be sharpened or re-moved to provide 2D images containing only in-focus features. These images canbe compiled into an informative 3D volume.This thesis develops some novel chemometric methods to improve the quan-titative and qualitative analysis of spectro/microscopic measurements. This studyuses vibrational spectroscopy, mainly Raman and NIR, to provide one-dimensionalsignals of pulp sample and consequently applies data mining techniques to processsignals and classify them to a degree that accurately predict the properties of prod-uct sheet. To study the data mining efficiency on two-dimensional data, we usedinterferometric Scattering microscopy to provide images of different types of bio-logical cell.1.1 Outline of This ThesisChapter 2 presents a brief overview of the instrumentational and computationalmethods used in this thesis. The instrumentational section contains an overall de-scription of the development and the concept of vibrational spectroscopy and inter-ferometric scattering microscopy. A discussion of advantages and disadvantagesof vibrational spectroscopy techniques and iSCAT are also provided. The compu-tational section includes one and two dimensional wavelet transform.Chapter 3 presents a fast computational method, Template Oriented GeneticAlgorithm (TOGA), that permits the accurate determination of a target chemicalspecies in the presence of interfering substances. We evaluate our feature selectionchemometric method on Raman spectra of a simple biological sample, containingdifferent monosaccharides.In chapter 4-6, we have investigated the effectiveness of Raman, NIR and the4combination of these techniques in classifying pulp spectra to predict the propertiesof product sheet in advance of its manufacture. We follow three goals: 1) A bet-ter understanding of the effect of morphology or chemical composition parametersin the physical- mechanical properties of paper sheet. 2) A better understandingof the influence of pulp processing parameters such as refining process on end-point product properties. 3) A practical technology to control paper-making pro-cess manufacturing with a degree of precision sufficient to enable novel productdevelopment.Chapter 4 addresses the two limitation factors for a full-scale application ofNIR spectroscopy as an on-line process control tool. These factors are (i) broadovertone absorption bands and (ii) unrelated variations in NIR spectrum to targetproperty. The first limitation makes it hard to discern a signature feature. The sec-ond limitation modulates the spectrum, and this tends to mask determinate spectralvariation. The present work explores the effectiveness of data processing strate-gies designed to remove uncorrelated variance from calibration models linkingNIR spectra with standard measures of paper quality, including tensile, tear, burststrength, wet and dry zero span length, freeness, absorption and scattering coef-ficients. This reports progress in overcoming the NIR limitations by utilizing thecombination of Orthogonal Signal Correction (OSC) and Discrete Wavelet Trans-form (DWT) signal processing technique.Chapter 5 describes the effectiveness of the combination of Raman spectroscopyand data mining techniques as an online technique in the pulp and paper industryto predict the final product properties. Raman spectroscopy offers a unique meansto rapidly measure pulp fibre properties and processing-related changes in theseproperties. However, characteristically weak signals, overwhelming fluorescencebackground, and matrix interference limit the application of Raman as on-line in-dustrial technique. Hence, we combine Raman spectroscopy with TOGA in aneffort to develop an appropriate process control tool for the pulp and paper in-dustry. This chapter analyses Raman spectra of unrefined pulp sample to predictstandard properties of paper made from that pulp at any arbitrary refining energy.The present study also investigates the effects of chemical bonds and the struc-ture of cellulose fibre on the mechanical properties of refined pulp using TOGAselected Raman signals.5Chapter 6 compares the effectiveness of NIR and Raman spectroscopy in pre-dicting pulp properties by applying the root mean square error of prediction toevaluate prediction model. Raman spectroscopy provides information with moredetails about the relationship between fibre chemical composition and sheet physi-cal properties. The NIR spectra in the other hand offer overlapping spectra limitedto overall information. Then, this chapter uses data fusion to determine whethercombining NIR and Raman spectral information provide prediction models withimproved efficiency in predicting physical paper properties. Chapter 7 investigatesthe performance of a new technique, which is a combination of interferometryscattering microscopy (iSCAT), confocal Raman spectroscopy, and deconvolutiontechniques in studying 1. The three-dimensional structure of different cells 2. Thedynamics of the interaction between a live cell and a gold nanoparticle when theparticle enters the cell, is in the cell membrane, and is in the cytoplasm 3. Thechemistry of the regions of the cell in which gold nanoparticles aggregate.This chapter provides a detailed description of the development and the conceptof iSCAT-Raman together with an extensive comparison with other microscopictechniques. It also describes in detail the applied deconvolution technique basedon the discrete wavelet transform.6Chapter 2Spectroscopy and Chemometrics2.1 Vibrational SpectroscopyVibrational spectroscopy relies on the changes of polarization or dipole momentduring molecular vibration. Two main spectroscopic techniques can detect changesin the vibrational state of a compound: Raman spectroscopy, which responds to thechanges in polarization, and infrared spectroscopy, which measures the changes indipole moment [71, 125, 298, 342]. The study of vibrational spectroscopy beganin 1800 when Sir William Herschel discovered infrared radiation with his well-known prism and thermometer experiment [255]. In 1930, the Indian physicistC.V. Raman won a Nobel Prize in physics for his observation in Raman effect[249, 282]. After the first fully automated IR spectrometer was built at 1940, theapplications of vibrational spectroscopy increased vastly.2.1.1 The Main PrincipleSpectroscopy is the science of the interaction of light and matter through the trans-fer of energy [73]. The energy of light depends on its frequency as follows:E = h.ϑ =h.cλ(2.1)Where h is Plank constant in m2.kgs , ϑ is frequency of light in s−1, c is speed of lightin ms , and λ is wavelength of the light in m.7Light in many regions of the electromagnetic spectrum excites atomic or molec-ular transitions corresponding in each case to a form of spectroscopy. Near-infraredspectroscopy refers to a wavelength range from 780 to 2500 nm while mid-infraredspectroscopy measures transitions between 2500 and 4000 nm.To develop a model for molecular vibrations, we start with the simplest possiblevibrating system, a diatomic oscillator. We can write the harmonic vibrationalstretching frequency in terms of a Hooks law force constant, K and for a diatomicmolecule:ϑ vib =12pic√kµ(2.2)reduces mass µ defined as:µ =m1m2m1+m2(2.3)c refers to the velocity of light, m1 and m2 are the mass of the two atoms.The force constant relates the strength of a chemical bond in a diatomic molecule.Equation 2.2 describes the effect of mass on the vibrational stretching frequency.The wavenumber of the absorption band ϑ is larger for stronger bonds and smallerreduced mass.Hooks law defines the potential energy of the harmonic oscillator (V ) as a func-tion of displacement of the atoms and force constant as follows:V =12k(r− re)2 (2.4)Where r is the internuclear distance between two atoms, and re is the internucleardistance between two atoms at equilibrium.The vibrational energy of a molecular harmonic oscillator is discrete as:Evib = hϑ(v+12)(2.5)where an integer v defines the vibrational quantum number.Quantum mechanics supplies an additional constraint for the harmonic oscil-lator, namely that only transitions between neighboring vibrational levels are al-8lowed.Figure 2.1 suggests that for a harmonic oscillator, a molecule can absorb in-finite energy without dissociation. However, in reality, all chemical bonds breakwhen enough energy applied to extend the vibrating bond. Figure 2.1 represents theanharmonic oscillator, which is a modification of harmonic oscillator to considerbond dissociation.  Dissociation Anharmonic Potential Harmonic Potential Figure 2.1: (left) harmonic model and (right) anharmonic model for the po-tential energy of a diatomic molecule.Equation 2.4 is modified to become the following equation by adding higherorder terms of displacement:V = kx2+ k2x3+ k3x4+ · · · (2.6)where k is force constant and k2 and k3 are much smaller than k.The Morse function uses an experimental equation to approximate an anhar-monic potential:V = De[1− e−a(r−re)]2(2.7)where a is a specific constant for each molecule, and De is the spectral dissociationenergy.Solution of the vibrational Schrodinger equation for a Morse function potential9yields:Evib = hϑ(v+12)− xmhϑ(v+ 12)2(2.8)xm is anharmonicity constant (0.005< xm < 0.05). The anharmonic model predictsthe occurrence of all transitions between adjacent and non-adjacent energy states.Figure 2.1 (b) shows in accord with equation 2.8 that energy difference decreasewith increasing quantum number.Vibrational spectroscopy provides information about chemical bond strength,the structure of a compound, the functional groups available in a compound, and afingerprint for almost every component [27, 47, 146, 245].2.1.2 Mid Infrared (Mid IR)A vibration is active when the electric field of incident light displaces atoms in avibrational mode , so that the dipole moment of the molecule changes. For funda-mental vibrational modes, each vibration must be active to appear in a spectrum.However, for the overtone and combination terms, only one of the combining vi-brations must cause a dipole change. The intensity of IR peaks is correlated withthe degree of anharmonicity and the magnitude of the dipole moment change whenincident light causes molecular vibrations [28, 142, 272].MIR absorption peaks represent the fundamental vibrational modes in a molecule.With the exception of homonuclear bonds, almost all chemical bonds appear inMIR spectra. Hence, MIR yields a fingerprint regime for various functional groupsin a molecule. An infrared spectrum yields information about length and strengthof the bond, and its conformational freedom [31, 150, 191, 233, 254, 260, 299].2.1.3 Reflectance Methods ( Attenuated Total ReflectanceSpectroscopy)Fahrenfort and Harrick introduced the concept of total internal reflection in the1960s [281]. When light at an incident angle of Θ propagates from a medium withrefractive index of n2 toward a medium with lower refractive index of n1 (n2 > n1),total internal reflection occurs at the interface of the two media, if Θ is larger thancritical angle (Θc) [86]. The critical angle is associated with the refractive indices10of the two media:θ c = sin−1(n2n1)(2.9)Attenuated Total Reflection (ATR) confines the IR beam in a dense ATR crystalthat is in contact with the sample. Since the refractive index of the crystal is high,total internal reflection occurs as illustrated in Figure 2.2.Figure 2.2: Schematic of a multiple reflection in the crystal of high refractiveindex in ATR instrumentAt each internal reflection, the superposition of the electric field of the incidentand reflected waves results in extending an evanescent field into the sample. Theproduced evanescent field inversely depends on the distance from the interface ofthe two media:E = E0e−[zdp](2.10)wheredp =λ2pi√n21sin2θ −n22(2.11)where z is the distance from the interface, E0 is the electric field at z=0 (interface),dp is the penetration depth, λ is the incident wavelength, θ is the incident angle,and n2 , n1 are the refractive indices of two media. Penetration depth describeshow far a fraction of the incident beam penetrates into the sample from the inter-face of sample and crystal. When the sample absorbs a specific wavelength, thisselectively causes energy loss in the beam [127, 151].The penetration depth is a small distance, so ATR can provide chemical-structure11information for highly absorptive or scattering samples [153, 313]. The main short-coming of this technique is its vulnerability to absorbing light with a lower fre-quency in the ATR crystal [360].2.1.4 Fourier Transform Infrared SpectrometerConventional IR spectrometry uses a grating to disperse light and select a wave-length, and only the spectral elements of one wavelength are captured by detector atany time. Interferometry techniques improve the speed of IR measurement due totheir capability for examining all wavelengths simultaneously. FTIR uses a Michel-son interferometer (Figure 2.3) to interfere two radiation beams with different path-lengths. The produced signal is a function of the variations of the path-lengths oftwo beams. Figure 2.3 schematically describes how a Michelson interferometerworks. A Michelson interferometer contains of one stationary and one movablemirror, which are placed perpendicularly to each other. An incident light beamenters the interferometer and a beam splitter lets half the intensity of light transmitto the fixed mirror, while the other half reflects to the movable mirror. The mov-ing mirror creates a difference in the light path, which causes a phase differencebetween two beams which recombine at the beam splitter.Figure 2.3: Schematic diagram of a Michelson interferometer used in FTIRtechnique.12After the two light beams are reflected by the mirrors through the beam splitter,they recombine and interfere. Based on the phase difference between two beams,they interfere constructively or destructively. The beam splitter again transmitshalf the intensity of light to the detector and reflects back the other half to thelight source. The result of FTIR instrument is an interferogram which shows thedetected light intensity versus the position of the movable mirror.The intensity of the detected signal (I (p,ϑ)dϑ ) is a function of variation inpath length (p) as follows:I (p,ϑ)dϑ = I (ϑ)(1+ cos2piϑ p)dϑ (2.12)The detected signal contains signals with different wavenumbers. The totalintensity I (p) Of the detected signal is the sum of all these signals:I (p) = (ϑ)(1+ cos2piϑ p)dϑ (2.13)Using the standard mathematical solution of Fourier transform, I (ϑ) can becalculated as:I (ϑ) = 4∫ [I (p)− 12I (0)](cos2piϑ p)d p (2.14)A computer, which is interfaced to the spectrometer, performs this transforma-tion [87, 110, 131, 161].2.1.5 Near Infrared (NIR)Near infrared spectroscopy is an efficient, non-destructive, and quick techniquethat requires no sample preparation. NIR produces complex and broad overlappingabsorption bands representing overtones and combination bands. As mentioned inthe last section, the dipole moment of a molecule should change in order to make itIR active. In addition to a dipole moment change, NIR requires a large mechanicalanharmonicity of the vibrating atoms. In the anharmonic model, vibrations are de-pendent to each other and can interact with each other. So, cross-terms from morethan one vibration are required to correct the total vibrational energy described inEquation 2.8 for an anharmonic system.13Evib =∑h.ϑr(vr +12)+∑∑h.xrs(vr +12)(vs+12)+ · · · (2.15)Here, ϑ r is the fundamental frequency and vr is the quantum number of vi-brational mode r. xrs is the anharmonicity constant for the interaction betweenvibrational modes r and s. NIR measures the interaction of vibrational modes pro-ducing overtone and combination terms. In overtone and combination terms, onlyone vibration must be active to make the interaction active in NIR, while in fun-damental mode all the vibrations must be active. So, some of the vibrations whichare not observable in the middle infrared can be studied in near infrared spectrum[35, 230, 288]. The S-H, O-H, N-H, and C-H bonds present high anharmonic-ity and high bond energy to produce fundamental vibrational transitions at higherwavelengths than 3000 nm. These two features allow us to measure the overtonesand combination tones of the bonds with NIR techniques [134, 136, 259, 338].2.1.6 InstrumentationA NIR spectrometer can use a visible-light optical component such as halogenlamp or tungsten coil as light source. Detectors based on silicon, PbS, and InGaAsphotoconductive materials are appropriate for the NIR region [199].Different NIR instrumentation use different techniques to produce monochro-matic spectra in the NIR region. Conventional NIR instruments disperse the lightusing diffraction gratings. To improve the scan speed over a broad NIR range,Acoustic-Optic Tunable Filter (AOTF) and Fourier- transform (using interferom-etry) based instruments [42, 185, 283] have been introduced. AOTF allows theconstruction of an NIR instrument without any moving parts. Instruments basedon acousto-optical tunable filters are capable of producing very high-speed scans,appropriate for online applications [111, 248, 351]. Usually more than one filter isrequired to identify a complex compound.Of all the NIR tools, Fourier transform NIR produces the best signal-to-noiseratio with a precise and accurate wavelength selection, however, it is slower thanAOTF spectrophotometers. Another modern technology in NIR instrumentation isusing LED sources with a band width of about 30-50 nm. LED sources are cheap14and many of them centered in different wavelengths can be used to cover all therequired wavelengths in NIR region [188, 222].2.1.7 Raman Spestroscopy and Its InstrumentationC.V. Raman and K.S. Krishnan were the first scientists who observed the phe-nomenon of inelastic scattering of light by an object in 1928. Two years later,Raman won a Nobel Prize in physics for his discovery of this effect, which wasnamed after him [164, 250].The interaction of incident light with electrons of a compound generates peri-odic vibrations in the electrons. Electron vibrations scatter light by producing anoscillating electric dipole moment. If the scattered light has the same frequencyas incident light, it is called Rayleigh scattering (elastic scattering). Scatteringproducing lower and higher frequencies than that of the incident light are knownas stokes and anti-stokes (both are inelastic scattering), respectively. When anexchange of energy between the incident light and the molecule happens, the scat-tered light gains an inelastic component, known as Raman scattering [108, 117].Figure 2.4: Three types of scattering by a molecule excited by a photon withenergy, E = hϑ .Figure 2.4 illustrates the three types of the scattering. Rayleigh scattering dom-inates the scattered light, while just 1 out of 106 of incident photons provides15Raman scattered light [234]. Raman spectroscopy usually suffers from high flu-orescence backgrounds with weak scattering signals. The low cross section of Ra-man scattering compared to fluorescence causes a large fluorescence background[119, 247]. Figure 2.5 shows an example of Raman spectra of fibre sample, show-ing Raman peaks on top of the large fluorescence background. Signal processingtechniques and some modifications in hardware, such as using an incident lightwith longer wavelengths, can mitigate the background.Figure 2.5: Raman spectra of fiber samples with large fluorescent back-ground.2.1.8 Resonance Raman SpectroscopyWhen it derives transition to a virtual state, beneath of an excited electronic state,the intensity of Raman scatter follows Albrecht-Hutley equation:IR ∝[(ϑL−ϑ vib)2(ϑ 2e +ϑ2L)(ϑ 2e−ϑ 2L)2]2(2.16)where, ϑL is the laser frequency, ϑ vib is the vibrational frequency, ϑ e is the elec-tronic transition frequency.16ϑL is usually much smaller than ϑ e, so equation 2.16 can be simplified to:IR ∝(ϑ l−ϑ vib)4ϑ 4e(2.17)Equation 2.17 shows that the intensity of Raman signal proportionally increasesby the fourth power of laser frequency. However, higher frequency raises the flu-orescence background. So, there is a trade-off between enhancing the signal andincreasing background.Figure 2.6: A simplified schematic diagram showing light absorption Stokesand anti-Stokes Resonance Raman scattering (RRS) processes.Figure 2.6 illustrates the mechanism of UV-Resonance Raman scattering. Whenthe laser frequency reaches the electronic frequency of the molecule, resonance Ra-man occurs if the laser frequency has crossed frequencies of vibrational states. Inthis condition, Raman scattering is enhanced by 3-5 orders of magnitude [17, 59,173, 212, 257, 297].2.1.9 Coherent Anti-Stokes Raman Spectroscopy (CARS)CARS is a non-linear Raman spectroscopy technique producing an enhanced anti-stokes Raman signal. CARS applies two coherent lasers to excite the sample. Oneof these has a constant frequency and the other has tunable frequency. As Figure2.7 represents, the difference between two laser frequencies should be chosen tobe equal to the frequency of the specific vibrational energy of the molecule. When17two lasers with fixed difference in their frequencies (∆ϑ l) interact with the sample,they produce a strong anti-stokes scattering signal (wCARS = 2wpump−wstokes) forthe vibrational mode with the same frequency as ∆ϑ l .Figure 2.7: Energy level diagrams describing, (left) spontaneous Raman scat-tering processes, (middle) stimulated Raman scattering (SRS), and(right) coherent anti-Stokes Raman scattering (CARS).The main shortcoming of CARS is a large background contributed from thematrix and other vibrational modes [32, 54, 203, 223, 258, 365].2.1.10 Stimulated Raman Scattering (SRS)In SRS, similar to CARS, two stokes and pump beams excite the sample. However,in SRS instead of two pump photons just one pump photon interacts with the stokesphoton.An SRS signal is generated when wSRS = wpump−wstokes matches a specificvibrational response based on the energy exchange between stokes and the pumplight beams. The intensity of the scattered light at the pump frequency loses energy(SRL: Stimulated Raman Loss), causing a gain in the intensity of the scatter lightat stokes frequency (SRG: Stimulated Raman Gain).The advantage of SRS over CARS is that SRL and SRG occur only if wSRS =wpump−wstokes matches a vibrational mode in the molecule. So, SRS is free fromnon-resonance background [95, 123, 186, 261, 330, 333].18SRG  SRL Stoke Pump Pump Stoke Vibrational State Figure 2.8: Diagram of stimulated Raman scattering principle. It describesthat SRS signal is generated when a specific vibrational response basedon the energy exchange between stokes and the pump light beams.2.1.11 Surface Enhanced Raman Spectroscopy (SERS)One of the most well-known and widely used techniques to enhance a weak Ra-man signal is SERS. According to the quantum mechanical expression (equation),Raman scattering intensity is proportional to the induced dipole moment:I = Nl64pi23c2(ϑ 0±∆ϑ)4|Plm| 2 (2.18)Where Nl is the number of molecules in the initial state, ϑ 0 is the frequency ofthe incident light, ∆ϑ is the Raman frequency shift, and Plm is the induced dipolemoment.On the other handPlm ∝ Eα0 (2.19)Where E is magnitude of the electric field, and α0 is the polarizability.Equation 2.19 explains that the induced dipole moment is proportional to Eand α0. Using equation 2.19 and 2.18 brings the conclusion that by increasing theelectric field, the Raman signal is enhanced. Using a rough surface of metal can19increase the electric field. Two mechanisms can explain this effect:1. The incident photons excite the conduction electrons located at the surfaceof the metal and cause a plasmon resonance. The created plasmon resonancepolarizes the surface and results in a large electromagnetic field.2. The metal and the adsorbed molecules interact in transferring charge, caus-ing an increase in the molecule polarizability.The most common nanostructure metals used for SERS are gold and silvernanoparticles. Gold and silver nanoparticles enhance the Raman scatter by theorder of 3-6 [213].2.1.12 Confocal RamanMarvin Minsky introduced confocal scanning microscope in 1957 [196]. In widefield Raman, a relatively large part of the sample is illuminated at once that scattersback a large amount of light. In confocal microscopy, the incident light illuminatesa single point and collects the scattered light from that single point. Figure 2.9illustrates that by setting a pinhole between the objective and the sample, a widefield microscope can be modified to become a confocal microscope. The pinholeblocks out all of the unfocused scattered light. Only the scattered light collectedfrom the confocal plane reaches the detector. This characteristic of confocal Ramangives it contrast and increases its ability to construct a 3D image [84, 158].20Figure 2.9: A schematic of conventional (left) and confocal (right) Ramaninstrument. In Confocal Raman just focused scattered light would bedetected by detector.21Table 2.1: Comparison of potentials, advantages and disadvantages of three different vibrational spectroscopic tech-niques (Raman, MIR, and NIR).Technique Raman MIR NIRBased on ScatteringAbsorption (basicvibration)Absorption (overtonesand combination tones)Activation conditionChange ofpolarizabilityChange of dipolemomentChange of dipolemomentSignal intensity Poor Good GoodBonds (mostapplicable) Homonuclear bonds Polar bonds H-containing bondsShape of signal Well- resolved Well- resolved OverlappedSelectivity High High LowParticle size Independent Dependent DependentInterferenceFluorescencebackgroundWater peak Water peakOnline technique Yes Usually not Yes222.2 Interferometric Scattering Microscopy (iSCAT)Single molecule microscopy has emerged as an important tool in an increasingnumber of complex material imaging applications [226]. Most of the single moleculedetection methods rely on Fluorescence. However, fluorescence suffers drawbacksincluding limited flux and the necessity in most cases to introduce fluorescent la-bels; the object of interest must be labeled with a fluorescent marker. Fluorescentmarkers are single quantum emitters, and are able to emit a certain number of pho-tons per unit time. The necessary condition for photon emission is the populationof an excited electronic state. Photochemical effects can limit the yield of emissionphotons. These limitations result in increasing the required observation time of asingle emitter to the order of a minute [79, 96, 204]. For a practical fluorescencemicroscopy, the achievable speed and localization of a single molecule roughlyfollow this equation:σ (time,space) = 1nmHz−12 (2.20)If high localization accuracy is desirable, imaging at high speed is ineffective.Therefore, this trade-off limits the applicability of fluorescence to studying thedynamics of many biosystems, where many processes occur at very fast rates [166,167, 276, 355].Extensive research has been carried out to develop an optical technique thatavoids the limitations of fluorescence [3, 19, 287, 295]. Scattering microscopyoffers a useful alternative, because detection and localization are limitted only tothe incident photon flux. Two scattering microscopies are of particular interest:pure scattering based, and interferometric techniques. Unlike in fluorescent mi-croscopy, the photon flux in scattering techniques is only restricted by the amountof incident power striking the object, and the scattering cross section of the object[118, 273, 289].2.2.1 Pure and Interferometric Scattering MicroscopyScattering microscopy can be considered as an alternative approach to fluorescencebecause only the incident photon flux limits detection and localization. We cantune the power of incident light to achieve a required scattered photon flux. This23advantage of scattering over fluorescence makes it theoretically possible to achieveunlimited temporal and spatial resolution. However, in a practical experiment,background scattering limits the signal to noise ratio, and consequently limits de-tection and localization.In scattering techniques, the number of photons observed by the detector varieswith incident light access to:N = Ni[ζ 2+ |s| 2+2ζ |s| cos∆φ](2.21)where s is the scattering amplitude, Ni is the number of photons in the incidentlight beam, ∆φ refers to the phase difference between the reference and scatteredelectric fields, and ζ refers to background (reference term).In the configuration of a pure scattering technique known as dark-field mi-croscopy, the only light detected is light scattered by the specimen, due to rejectionof illumination light coming from the background. So, the detected signal for purescattering depends for the most part on the term |s| 2, while for interferometric scat-tering, the reference term ζ 2 dominates the signal. Therefore, we can simplify theabove equation for the two scattering approaches as follows:Dark–field:N ≈ Ni|s| 2 (2.22)Interferometric Scattering:N ≈ Ni[ζ 2+2ζ |s| cos∆φ](2.23)The background term has no information about the sample, so that the inter-ference term is the only contribution includes information of the specimen. Thus,the signal produced by the sample appears as a small variation on top of the largebackground. The ratio of the signal term to the background term, in other wordsthe signal contrast, is described as:contrast =IsIb(2.24)24For pure scattering:contrast =∣∣s2∣∣ζ(2.25)For interferometric scattering:contrast = 2|s|cos∆φζ(2.26)For objects with diameters smaller than the wavelength of the incident light,the scattering amplitude depends on the size of the object as follows:s = γεm (λ )piD32λ 2ε p (λ )− εm (λ )ε p (λ )+2εm (λ )(2.27)where s relates to dielectric constant of the particle, ε p , and the medium εm. γ is aproportionality constant related to polarizability, and λ is the incident wavelength.Equations 2.25 and 2.26 show that for interferometric scattering, the signal con-trast is linearly proportional to the scattering amplitude, and consequently based on2.27 signal contrast is proportional to the particle diameter cubed. However, theproportionality of signal contrast to particle size increases to the sixth power forpure scattering.This makes a big difference between interferometric scattering and either dark-field or fluorescence. While fluorescence and dark-field typically have zero back-ground, interferometric scattering deals with a large background to provide goodsignal contrast.The limitation of pure scattering-based techniques is their high dependencyon the size of the object. Interferometric scattering improves this limitation anddecreases the order of the relation of the size with the intensity. However, in in-terferometric scattering, the background signal typically overwhelms any signalsderived from sample features. Fortunately, the background signal is almost con-stant and can be removed by differential imaging [140, 237, 364].2.2.2 Disadvantages of Scattering MethodologiesThe previous section discussed the limitations of fluorescence and how scatteringapproaches can overcome these shortcomings. However, scattering microscopy25suffers from some disadvantages such as:• Scattering is non-specific in the detection of single objects, information read-ily available with fluorescence.• Scattering is strongly dependent on object size.• Scattering suffers from high background noise and spurious reflections.• Scattering efficiency is dependent on the interaction of the incident light witha single object, which is wavelength dependent.For much smaller particles compared to the wavelength of light, the detectedscattered light is given byE = s.Ei = s.Eiexp(iφ) (2.28)E = sEi = sEi exp(iφ) (2.29)Using general Mie theory to explain the scattering of all objects with mis-matched refractive indices with their surroundings, the scattering amplitude can begiven by:σ = |s| 2 ∝V 2∣∣∣∣∣n2p−n2mn2p+n2m∣∣∣∣∣2(2.30)Here, np and nm denotes the refractive indices of the object and the medium,respectively. The scattering amplitude depends on the refractive indices of mediumand object, which are wavelength dependent [18, 38, 67, 140, 198, 209, 237, 364].2.2.3 Using Gold Nanoparticles to Address the Size Dependency ofScattering TechniquesGold nanoparticles scatter visible light strongly owing a surface plasmon resonancephenomena. At the frequency of visible light around 500 nm, gold nanoparticleshave a high polarizability, which causes an enhanced scattering amplitude. Thisallows the observation of very small particles with the same scattered photon flux.26For example, a 40 nm gold nanoparticle in water scatters light with an amplituderoughly equal to that of a 160 nm diameter of silica bead [140, 176].2.2.4 Addressing the High Dependency of Dark-Field on Particle SizeIn this section, we introduce interferometric scattering as a method developed tominimize the limitations of pure scattering microscopy. As discussed in the lastsection, dark-field microscopy removes the background light whereas interfero-metric scattering collects both background light and object scattered light. As men-tioned before, the detected light in interferometric scattering relates to the incidentfield, Ei, by:Idet = |Er +Es| 2 = |Ei| 2[ζ 2+ |s| 2+2ζ |s| cos∆φ](2.31)where ∆φ is the phase difference between scattered and reflected electric fields.For a glass-water interface, the term ζ is roughly equal to 0.065 while the sterm for a particle smaller than 50 nm is much smaller than 0.065. Comparing thesignal contrast of interferometric scattering with dark-field scattering shows thatthe contrast scales linearly with scattering amplitude while in dark-field it scaleswith scattering amplitude squared. Consequently, instead of depending on D6 asin dark-field, interferometric scattering depends on D3. This decrease in the orderof size dependency makes interferometric scattering able to detect an object assmall as 5 nm. This size limit is an order of magnitude smaller than the size of thesmallest detectable object by pure scattering [15, 97, 167, 208, 291].2.2.5 History of Interferometric ScatteringCurtis introduced first interferometric scattering detection in 1965 with a techniquecalled interference reflection microscopy (IRM) [67]. Ploem developed this tech-nique by introducing anti-flex illumination scheme [239]. Kukura and Sandoghdarused a coherent light source to improve the contrast of iSCAT compared to theaforementioned interferometric techniques [229].Most reported applications of iSCAT track the lateral position of a specimenin a large heterogeneous system. This is well suited to biological applications,to develop transport models for different biological samples. Ortega-Arroyo and27coworkers attached gold nanoparticle to the object of interest, such as single lipid,to study its dynamic motion [15, 28, 97].Most of the studies utilizing iSCAT have emphasized the use of nanoscopiclabels; however, iSCAT applications can extend to any molecule of interest. Whenthe refractive indices of an object and its surroundings are mismatched, the objectcan scatter a portion of the incident light, making it detectable by iSCAT.2.2.6 Combined iSCAT and Confocal Raman MicroscopyAs mentioned in the last section, iSCAT collects back-scattered light from the sam-ple and compares the phase information to a reference field to create an image. Theresult is a wide field view of a label-free live sample capable of video frame rates inreal time. However, iSCAT images do not contain any chemical information apartfrom the refraction index difference. On the other hand, Raman spectroscopy isa powerful method for label-free imaging and for gathering chemical informationabout a sample, but is limited by long acquisition times. In a Raman mapping ex-periment, it is important to validate that the system being studied does not changedramatically on the timescale of the measurement owing especially to the incidentlaser excitation energy. The shortcomings of Raman makes Interferometric scat-tering microscopy (iSCAT) a useful complementary imaging technique. The com-bination of these two techniques provides a real-time image of a sample togetherwith the Raman spectrum of the region of interest.A continuous wave 532 nm diode-pumped solid state (DPSS) laser, and aHelium-Neon laser emitting at 632.8 nm, for iSCAT microscopy and confocal Ra-man microscopy, respectively. Figure 2.10 depicts the experiment setup of iSCAT-Raman. All the optics are mounted on an isolated optical table to minimize thevibrational effects on the measurement. A combined piezoelectric and mechani-cal translation stage moves the sample in x, y, and z -directions relative to a fixedobjective.We use two perpendicular Acousto-Optic beam Deflector (AOD) crystals toraster the iSCAT illumination beam across the entrance aperture of the objectivein a rectangular shape, at a fixed frequency. The selected frequency is much fasterthan the exposure time of the CMOS detector. This allows us to obtain wide-field28images. In our iSCAT-Raman set up, both X and Y AODs with model (45070-5-6.5DEG-.63) by Gooch and Housegeo operate that the frequency between 50and 90 MHz provided by two Voltage-Controlled Oscillators (VCO). An OlympusPLAPON 60X0 infinity-corrected oil-immersion objective with the high numericalaperture of 1.42 focuses the iSCAT and Raman beams onto the sample. However,the magnification of this objective is 60x, the higher numerical aperture yieldsfocus with high light collection efficiency forms a tight diffraction limit .A high-speed CMOS (PtGrey- model GS3-U3-41C6M) camera was chosen asthe iSCAT detector due to its high achievable readout rate of about 22 microsec-onds. Depending on the application of iSCAT we are using, we adjust the pixelresolution of camera. For most applications, a resolution of 80-100 nm per pixel,with a pixel-to-pixel signal variation of 0.1 % , would be desirable. In theory, withthe assumption of a shot noise limited measurement, the collection of 106 photonsper pixel provides this 0.1 % variation. However, most CMOS sensors have wellcapacities in the range of tens to hundred thousands of photoelectrons. To deal withthis limitation, we bin 9 neighboring pixels together, yielding a super pixel, withthe number of photoelectrons close to 106. This approach requires triple magnifi-cation, leading to the pixel resolution of 33 nm. In our setup, we used an imaginglens to achieve the necessary magnification.A monochromator (IsoPlane SCT 320) with an attached CCD camera (PIXIS100BX) was served as a spectrometer for Raman. The CCD has an imaging areaof 26.8 by 2.0 nm with an accuracy about 0.1 nm. The high-NA oil immersionobjective focuses laser light on sample and collects scattered signal. This objectiveis apochromatic to correct both chromatic and spherical aberration [57].29Figure 2.10: Schematic diagram of the combined iSCAT-Raman microscope, including lasers operating at 532 nm(iSCAT) and 633 nm (Raman), acousto-optic beam deflectors (AOBD), long-pass filters (l p f ), CMOS camerasfor iSCAT and Bright-field, confocal apparatus, and spectrograph. Note that the bright-field system is peripheraland was not used in the present research.302.2.7 Limitations of iSCATIn theory, the only factors limit the quality of an image and the rate of imageacquisition by iSCAT are shot noise and the frame rate of the camera, respectively.The signal to noise ratio of an image depends linearly on the root of the number ofphotoelectrons detected by the camera and on the signal contrast according to thefollowing equation:SNR = S∗N −12 (2.32)An increase in laser power improves N and consequently localization precision.In reality, however, background scattering limits the SNR and localization. Dif-ferent factors such as reflection from dust and the aperture (such as lenses in theobjective), refractive index changes at different depths of a sample, and limitedcapability of objective in collecting light, all deteriorate the expected theoreticalSNR [229].2.3 Factors Affect the Quality of Images2.3.1 Point Spread Function ModeWhen light is emitted from a small source, the objective is not able to focus lightonto an extremely small point in the image plane. Because light waves convergingon the focal plane interfere to yield a diffraction pattern of light. The pattern con-tains concentric rings of light surrounding a bright disk in the center. The disk iscalled an Airy disk, and its size depends on the numerical aperture of the objective.For a perfect optical system with no spherical aberrations, the diffraction patternlooks symmetric and periodic in the axial and lateral planes. Depending on thetype of microscopy being used, the diffraction pattern can take different shapes.Hourglass or football shapes are the most common diffraction patterns in the axialdirection. The axial shape of the point spread function (diffraction pattern) limitsthe ability of system to resolve two objects which are close to each other. Howclose two objects can be while still being resolved is known as axial resolution,defined as:31raxial =2λnNA2(2.33)Here, NA denotes the numerical aperture, n the refractive index of the medium,and a given wavelength. The two dimensional point spread function of an aberration-free optic can be described as:h(ϑ) = 2∫ 21P(ρ)J0 (ϑ ,ρ)ρ dρ (2.34)whichρ =√x2+ y2a(2.35)andϑ =2piNAλ√x2+ y2 (2.36)Here, P denotes the circular pupil function of radius a, λ the emission wave-length, and J0 the first order Bessel function. Equation illustrates that point spreadfunction depends on emission wavelength and numerical aperture [77, 239, 294].2.3.2 Aberration in Image FormationMost objects, especially biological specimens, have different refraction indices indifferent layers of the object. When light passes through these kinds of the objects,optical paths diverge from the optimum path for the objective, causing sphericalaberration. The only Z plane with no spherical aberration is the one just above thecoverslip. Spherical aberration affects the shape of the axial point spread function.2.3.3 Chromatic AberrationTwo factors cause chromatic aberration: a) the light source contains multiple wave-lengths, i.e. it is not a line source. An objective does not focus all the wavelengthsin exactly the same spot; different wavelengths have different optical paths. B)Different wavelengths have different refractive indices, leading to different focallengths. The refractive index increases by decreasing the wavelength. Using amonochromatic light source is the best correction for this aberration. Achromatic32Chromatic AberationFigure 2.11: Schematic diagram of chromatic aberration of a single lenslenses are also able to reduce chromatic aberration [163, 357].2.3.4 Spherical AberrationThe geometry of a lens causes spherical aberration, because of the higher refractionof light at the periphery (near the edge of the lens) rather than the central portion.This difference in refraction causes light rays focus at different points.Spherical AberationFigure 2.12: Schematic diagram of spherical aberration of a single lensTo minimize this aberration, using an aperture to restrict incident light to thecentre of the objective is common. Aspheric lenses can also be used to reducespherical aberration [154, 263, 279].2.3.5 COMA AberrationDifferent lens zones have different magnifications, which causes an axial -cometlike image at some distance from the principal axis of the system.33Lenses with wider apertures cause greater COMA aberrations. So, this aberra-tion can be minimized by reducing aperture size [279].Figure 2.13: Schematic diagram of coma aberration of a single lens2.3.6 AstigmatismRays that travel along perpendicular planes have different focal points, leading toimage stretching in one direction at one focal plane, and an opposite direction atanother plane. Astigmatism reduces the sharpness of an image.TangentialSogitalOpical AxisFigure 2.14: Schematic diagram of astigmatism effectImperfect lens surfaces cause astigmatism. So, using lenses with perfect sur-faces decreases astigmatism [353].342.3.7 Noise in Image FormationTwo features of spectroscopy produce noise: the quantization of light as photons,and the conversion of photoelectron charges to a pixel. Four main types of noisedeteriorate the quality of an image including: thermal noise, shot noise, dark cur-rent noise, and readout noise. iSCAT images suffer mostly from shot noise, withrandom-count characteristics.2.3.8 Shot NoiseThe detector receives photons and converts them to photoelectrons. The sponta-neous rate of conversion corresponds with the incident power at the detector, asfollows:α (t) =η (λ )hνW (t) (2.37)Here, η (λ ) denotes the detector quantum efficiency, α (t) the instantaneousconversion rate, W (t) the instantaneous incident power at the detector. Equation2.37 shows that instantaneous rate is proportional to the instantaneous incidentpower. Photons striking the detector occurs at random intervals, making W (t) astochastic parameter. Therefore, the number of photons striking the detector andconsequently the number of generated photoelectrons varies during the detectorintegration time. These fluctuations generate shot noise in an image. Shot noisedegrades the signal to noise ratio and the contrast of an image. The exposure timeand light intensity affect shot noise in a way such that lower exposure time andlight intensity decreases the image contrast. The root-mean-square of shot noiseis proportional to the square root of the image intensity. Shot noise per pixel isindependent from other pixels and follows a Poisson distribution [78, 145, 280].2.3.9 Thermal noise (Johnson-Nyquist Noise)Thermal noise is a Gaussian additive noise which is independent at each pixel andindependent of the signal intensity. The finite temperature of the charge carrierscauses a small current to flow in a circuit. Thermal noise is a random noise andfalls in a Gaussian distribution over time.352.3.10 Dark Current NoiseThe thermal energy in detector creates electrons over time that are not produced bylight striking the detector. The detector captures these electrons and counts themas signal. Thus, even in the absence of light, a signal can be detected due to thethermally generated electrons. The rate of thermally generated electrons increasesexponentially with temperature. The dark noise has a similar characteristic to shotnoise: both are sources of Poisson noise. Averaging several dark frames togetherreduces the dark current noise.2.3.11 Readout NoiseDuring the photoelectron digitization process, readout noise arises due to the lackof capability of the amplifier to measure the charge in the clump of electrons per-fectly. There is a random fluctuation in the amplified readout. It is independent onthe exposure time and the intensity of the detected photons.2.4 Wavelet Transfrom2.4.1 Fourier Transform (FT)The Fourier Transform employs an infinite summation of sines and cosine func-tions, as basic functions, to extract dominant underlying frequencies in a sig-nal. Given the embedded trigonometric waveform properties in its construct, theFourier Transform is well suited to study stationary signals where there could bean infinite appearance of all frequencies. Hence, the Fourier Transform assumesthat the underlying information is global which makes it less applicable to signalswith localized information. The transformation of a time-domain single-value real-variable function f(t) to a frequency-domain single-value complex-variable func-tion F(w) via the Fourier Transform is given byF (w) =∫ +∞−∞f (t)e− jwtdt =∫ +∞−∞f (t) [cos(wt)− jsin(wt)]dt (2.38)36where j is the complex number. The integration procedure involved in the FourierTransform is commonly known as the inner product of f(t) and e− jwt , which isbasically a measure of the similarity of two functions; the higher the value of theintegral, the higher the similarity (local matching) between the two functions. So,the integration procedure serves to decompose the original signal f(t) to a linearcombination of trigonometric functions of different frequencies. The outcome ofthe inner product for a given frequency w0 is called the Fourier Coefficient corre-sponding to the frequency w0. If the original signal contains significant informationaround the frequency w0, then the plot of |F(w)|, the magnitude of complex func-tion F(w), in the frequency domain contains a large local value at w0 compared to|F(w)| at other frequencies.To illustrate this procedure, the signal f (t) = 0.6 ∗ sin(2 ∗ pi∗ 30 ∗ t)+ sin(2 ∗pi ∗ 80 ∗ t), and X = S+ 2 ∗ randn(size(t)) as white noise; which is composed oftwo frequencies f1 = 30Hz and f2 = 80Hz, is superposed with a random function,shown in Figure 2.15(a). The original signal underlying frequencies are no longeridentifiable in the new signal due to the existence of the background noise. Thepower spectral density function of the discrete Fourier Transform of this signal isable to reveal the underlying frequencies, shown in Figure 2.15(b).Figure 2.15: Illustration of the decomposition of an apparently random signal(left) to its principal underlying frequencies at f1 = 30 Hz and f2 = 80Hz (right).As stated above, the single-variable complex-valued Fourier Coefficients existspurely in the frequency domain. There are a large number of applications, such as37image, voice, in which information is distributed both in time and frequency. Thewindowed Fourier Transform (aka short-time Fourier Transform) and the wavelettransform are constructs similar to the Fourier Transforms aiming at informationcontaining both time and frequency. These are the subject of the following sections[48, 76, 175].2.4.2 Short-Time Fourier Transform (STFT)All Fourier coefficients in equation 2.38 are obtained by integration over time andas a result are time independent subject to stationary analysis. In reality, a signalmight last for only a short interval or could occur at irregular intervals. The FourierTransform can still provide the frequency content of a non-stationary signal, how-ever, is not able to provide useful information about the duration of the signal orthe beginning and end of an intermittent signal. These deficiencies of the stationaryFourier Transform analysis are removed by the non-stationary Short-Time FourierTransform, where the transform is a function of both time and frequency. To do so,a time function also known as a window function is considered and shifted in timeby to center around t=τ . The original Fourier Transform, equation 2.38, is nowwritten asF (w,τ) =∫ +∞−∞f (t)g(t− τ)e− jwtdt (2.39)The window function can be considered as a temporal weight function for theoriginal function f(t). The window function, lives on a short internal centeredaround t=τ , construct a proper time resolution within the original signal. Thesmaller the window size, the higher the time-resolution. Clearly, there is a com-promise between time-content and frequency-content of the original signal, knownas uncertainty relationships between the time and frequency support of a signal:the broader support in time (lower resolution) results in a narrower support (higherresolutions) in frequency, alternatively, a smaller window size improves time reso-lution while resulting in lower resolution in frequency . This shortcoming of STFTis relatively resolved by the Continuous Wavelet Transform (CTW), described inthe next section, by incorporating variable-sized windows in its formulation; longtime intervals are used when low-frequency information is required while short38time intervals are used for regions with high-frequency information [9, 109, 216].2.4.3 Continuous Wavelet Transform (CWT)The Continuous Wavelet Transform similar to FT and STFT employs the innerproduct of the original signal with a multivariable weight function. One may con-sider the temporal window function of STFT in equation 2.39 as the weight func-tion for the complex exponential e(− jwt) function and re-write equation 2.39 asF (w,τ) =∫ +∞−∞f (t)kτ,wdt (2.40)wherekτ,w = g(t− τ)e− jwt (2.41)which can be replaced by the followingkb,a(t) =1√aΨ∗(t−ba)(2.42)to construct the Continuous Wavelet Transform (CWT) asCWT (a,b)=CWT (a,b; f ,ψ)=∫ +∞−∞f (t)1√aΨ∗(t−ba)dt =∫ +∞−∞f (t)ψ∗a,b(t)dt(2.43)where a is the window size or the dilation parameter and b is the location parameterdenoting to the translation in time. The asterisk indicates that the complex conju-gate of ψ , the mother wavelet, is used in the transformation. An increase in thedilation parameter expands the wavelet in time (lengthening of time periods) whilecontracting its frequency spectrum; the dilation parameter is inversely proportionalto all the characteristic frequencies. Hence, to detect coarser features of a signalby wavelet coefficients, a more stretched wavelet is required. Clearly the CWTcoefficients are dependent on the choice of the mother wavelet. The choice of thewavelet usually depends on desirable features of the original signal to be detectedby the CWT.Figure 2.16 shows some wavelets that are used in the following chapters ofthis thesis. The wavelet shown in Figure 2.16 (a) is appropriate for a signal con-39taining abrupt discontinuities, while the wavelet in Figure 2.16 (b) is for a signalpossessing smooth oscillations [68, 81, 328].    Ψ(X)  X  (a)    Ψ(X)  X  (b)Figure 2.16: Examples of mother function of wavelet suited for signal witha) abrupt discontinuities , b) smooth oscillationsTo recover the original signal from its wavelet transform, the CWT are inte-grated over all scales and locations as followsf (t) =1cg∫ +∞−∞CWT (a,b)ψ ta,bda.dba2(2.44)where cg is the admissibility constant defined ascg =∫ pi0|ψˆ(w)|2wdw (2.45)and ψˆ(w) is the Fourier Transform of ψ(t), i.e.,ψˆ(w) =∫ +∞−∞ψ(t)e− jwtdt (2.46)For example, the value of Cg for the Mexican hat (t) = (1− t2)e−t22 , the secondderivative of the Gaussian function e−t22 , is equal to pi . Note that the total energy ofa wavelet could be obtained in the time-domain or the frequency domain asE =∫ +∞−∞|ψ(t)|2 dt =∫ +∞−∞|ψˆ(w)|2 dw (2.47)402.4.4 Discrete Wavelet Transform (DWT)The CWT discussed in the previous section was a two-parameter representation ofa continuous function f(t). However, data in many applications are represented bya finite number of values, where it is possible to reconstruct the data using infinitesummations of discrete wavelet coefficients. To this end, the scale and locationparameters are first discretized asa = am0 ,b = nb0am0 (2.48)where the integers m and n are, respectively, the control parameters for dilation(scale index) and translation (location index). a0 > 1 and b0 > 0 are, respectively,a fixed dilation step parameter and a fixed location parameter. The discrete wavelettransform cm,n , known as detail coefficients, using the discrete wavelets is thencm,n =∫ +∞−∞f (t)a√am0ψ(t−nb0am0am0)dt =∫ +∞−∞f (t)ψm,n(t)dt (2.49)and the original function f(t) can be reconstructed using the inverse discrete wavelettransform as followsf (t) =m∑m=−∞∞∑n=−∞cm,nψm,n(t) (2.50)given the wavelets are orthogonal to each other and normalized to have unit energy,i.e.∫ +∞−∞ψm,n(t)ψm′,n′(t)dt = 1, i f m = n and m′ = n′0, otherwise (2.51)The choices of a0 = 2 and b0 = 1 for the wavelet discretization is referred to dyadicgrid arrangement giving the following scaling function (or orthogonal dyadic dis-crete wavelet)Φm,n(t) =1√2mφ(2−mt−n) (2.52)41Figure 2.17: Multiresolution decomposition of a Raman signal of a pulp sam-ple using symlet 5 as wavelet function2.4.5 Two-Dimensional Wavelet TransformMany applications produce and display information in two-dimensional arrays, re-quiring a two-dimensional discrete wavelet transforms and algorithms. One of themost commonly used arrangements, termed a separate transforms, use the two-dimensional scaling function compared of the same scaling function in the hori-zontal and vertical directions. Thus for a two-dimensional scaling functionφ(x,y) = φ(x)φ(y) (2.53)the detail function is obtained from the following two dimensional horizontal, ver-tical and diagonal wavelets: (horizontal)φ(x,y)h = φ(x)ψ(y) (2.54)(vertical)ψ(x,y)ϑ = ψ(x)φ(y) (2.55)42(diagonal)ψ(x,y)d = ψ(x)ψ(y) (2.56)where x and y represent spatial Cartesian coordinates. The multiresolution decom-position of the two-dimensional coefficient matrices can be expressed asSm+1,(n1,n2) =12∑k1∑k2ck1ck2Sm,(2n1+k1,2n2+k2) (2.57)T hm+1,(n1,n2) =12∑k1∑k2bk1ck2Sm,(2n1+k1,2n2+k2) (2.58)T vm+1,(n1,n2) =12∑k1∑k2ck1bk2Sm,(2n1+k1,2n2+k2) (2.59)T dm+1,(n1,n2) =12∑k1∑k2bk1bk2Sm,(2n1+k1,2n2+k2) (2.60)where k1, k2 are the scaling coefficient indices and n1, n2 are the locationindices at level m+1 [189, 303].43Chapter 3Template-Oriented GeneticAlgorithm Feature Selection ofAnalyte Wavelets in the RamanSpectrum of a Complex MixtureWe introduce a fast computational method for feature selection that facilitates theaccurate spectral analysis of a chemical species of interest in the presence of over-lapping uncorrelated variance. Using a genetic algorithm in a data-driven approach,this method assigns predictors according to a template determined to minimizeprediction variance in a calibration space. Termed, Template Oriented GeneticAlgorithm (TOGA), this method establishes features of greatest significance anddetermines their optimal combination. We demonstrate the efficacy of TOGA fora set of model systems in which we seek to quantify a target monosaccharide inmixtures containing other sugars added in random amounts. The results establishTOGA as an effective and reliable technique for isolating the spectra of specifiedsubstances in complex mixtures.443.1 IntroductionSaccharides and polysaccharides (glycans) represent a large class of organic sub-stances in which small variations in structure give rise to an enormous functionaldiversity. It remains an important challenge in glycobiology to find efficient ana-lytical techniques by which to structurally elucidate saccharide and polysaccharideisomers with similar chemical and physical properties. Exacting methods of gly-can detection often involve wet-chemical steps of derivitization, separation andlinkage analysis in combination with mass spectrometry and NMR spectroscopy[215]. However, procedures such as these come at a cost associated with the needfor large samples and long analysis time by skilled personnel.Non-destructive spectroscopic probes can provide information of utility forbroader classification or screening purposes. Among such techniques, Raman spec-troscopy offers advantages for the analysis of biological substances. It requires lit-tle or no sample preparation and collects a relatively small background signal fromwater [33]. Raman calibration models accurately predict glucose concentrationsin pure water to physiological levels [82]. With a sufficient data base of authenticstandards, one could imagine Raman assays for polysaccharide identification andmonosaccharide quantification [214, 275, 337, 359].However, most biological substances apart from water exhibit about the samemass-equivalent spontaneous Raman scattering response, and this uniform inten-sity of signal causes difficulties for the analysis of complex materials. Samples ofmixed composition present overlapping Raman bands, creating a need for discern-ing analysisConventional spectrometric determinations often default to linear regressionmodels that refer to the relative intensity of a single feature. Such classificationmodels that rely on univariate analysis suffer most from spectral overlap. Sub-stances in a mixture signal their presence by signals at different Raman scatteredwavelengths, and measurements geared to a multivariate analysis of the spectrumas a whole yield far better results for mixtures of varying composition.Biological systems often present natural variations that bear no relationship to adetermination of interest. Such variations give rise to Raman spectral features thatchange in a manner that does not correlate with changes in an analyte or property45targeted for classification. Such uncorrelated variations degrade the quality of acalibration model. For cases in which uncorrelated variations appear in spectralregions apart from structure signifying the target analyte or property, selecting anappropriate subset of the most correlated features yields a more accurate, morerobust multivariate calibration model.However, feature selection carries with it the risk of losing relevant informa-tion. Chemometric approaches enable the analyst to make such decisions withgreater reliability [85, 144, 165]. Standard techniques include, iterative variableselection (IVS )[177], iterative predictor weighting (IPW)[50, 183], uninforma-tive variable elimination (UVE)[43, 46], GOLPE [30, 60], and interactive stepwiseelimination (ISE)[92, 93], stepwise selection forward (SVS) [56], or eliminationbackward (BVE) [180, 304], and stepwise interval partial least-squares (backwardor forward:BiPLS or FiPLS) [200, 225] .These methods represent modelwise elimination approaches [93]. That is tosay, instead of selecting a subset of features, (e.g. intensity values in the space ofRaman wavelengths), these methods offer ways to eliminate useless predictors inthe calibration algorithm; they exclude useless features according to their coeffi-cients in a regression. In so doing, these methodologies make strong assumptionsabout the spectral parameters, such as peak widths, and this can give rise to prob-lems when the character of the fitness space is unknown.The use of a genetic algorithm provides a means to deal with this limitation.This approach applies the principle of natural selection to choose features withoutmaking any assumption about the search space [171, 197, 220]. Genetic algo-rithms differ from the conventional feature-selection methods in several importantways. The methods mentioned above select single variables scattered through-out the spectrum, with the assumption that no correlation exists among variables.Variables selected by a genetic algorithm tend to concentrate in spectral regions ofgreatest relevance [107, 170]. Genetic algorithms use payoff information, definedby a fitness function, instead of derivatives or other auxiliary knowledge. In ad-dition, genetic algorithms use probabilistic transition rules, whereas most methodsuse deterministic ones [181, 182].The present study introduces a variant of genetic algorithm feature selectionadapted specifically to the problem of optimizing the prediction accuracy of a spec-46trochemical classification model. Our method uses a data driven approach, in effectconstructing an information template from the available data to guide the processof selection. We refer to this as a Template Oriented Genetic Algorithm (TOGA).In practice, TOGA first determines the significance of each feature and thenfinds an optimal combination of features. Here, the data itself controls the flowof information, not a predetermined program logic. Instead of randomly select-ing predictors, relying on event probability as in a conventional genetic algorithm,TOGA assigns predictors with weights defined to minimize prediction variance.We apply this data-driven approach using a wavelet transform representation ofthe spectrum in order to isolate realistic spectral features over ranges of wavelengthas opposed to picking features as individual intensity values at discrete wavelengths[26]. As a consequence, TOGA does not require pre-defined, target domain knowl-edge to isolate features that connect meaningfully to chemical information.In sections below, we show how one can use the variance as a guide to deter-mine an optimum feature-selection filter. We show how the application of this filterin building a calibration model vastly improves classification, and how the filteredfeatures point to the true spectrum of the variance that underlines a raw data setobscured by uncorrelated contaminants.To provide a controlled demonstration of TOGA, we adopt a system of monosac-charides as a model. Testing for prediction accuracy, we analyze for a targetmonosaccharide in a set of solutions prepared to exhibit an overwhelming uncorre-lated variance produced by random amounts of other monosaccharides. The resultsobtained for our well defined set of complex materials shows how TOGA can beused to advantage to quantify a substance or classify a property in the presence oflarge, overlapping uncorrelated variance, and provide an isolated representation ofthe spectrum that most relates to that classification.3.2 Materials and Methods3.2.1 SamplesStarting with single-component stock solutions of glucose, galactose, mannose andarabinose, each dissolved at 0.1 g ml−1 in deionized water, we performed serial47dilutions to fashion 20 pure standard solutions of each sugar with concentrationsin the range from 0.04 to 0.1 g ml−1 (c.f. Figure 3.1). We used each of thesesolutions to make three more standards at each analyte concentration. To each ofthese additional standards, we added random amounts of the other three sugars.By this procedure, we formed a sample universe of 320 solutions, consisting of 80pure sugar solutions, together with 240 solutions paring each pure sugar at eachconcentration with three different random amounts of the other sugars.The mix with other sugars serves to create an uncontrolled matrix of com-pounds with chemical structures similar to the compound of interest, with the pur-pose of producing an uncorrelated Raman background of similar signals. For vali-dation, we separately prepared five independent solutions for each sugar containinga known concentration of the target analyte, together with random amounts of theother sugars.400 600 800 1000 1200 1400 1600 1800 2000Frequency (cm−1)Student Version of MATLABFigure 3.1: (left) α-D-pyranose structures of monosaccharides, glucose,galactose, mannose and arabinose. (right) Raw Raman spectra of puremonosaccharides in aqueous solution at concentrations of 0.03 g ml−1,from the bottom, glucose, galactose, mannose and arabinose.483.2.2 Spectroscopic Instrumentation and MeasurementWe recorded Raman spectra of each sample using a Raman micro-spectrometer thatfiibre-couples an IPS 300 mW 785 nm diode laser to an Olympus BX-51 micro-scope. This system collects backscattered light with a six-around-one fibre bundle.The bundle exit, arranged as a vertical fibre stack, forms a 100 µm slit at the en-trance of an Acton 300 mm monochromator coupled to a Princeton instrumentsPIXIS back-illuminated CCD detector. A laboratory computer reads the CCD, andbins the pixels in each column, t, from 1 to 1400 to form a spectrum of scattered-light intensity versus Raman shifted wavelength.We prepare samples for analysis by filling a 100 µl aluminum well plate. Spec-tral acquisitions combine the signals recorded from five accumulations using a 5 sexposure time for each. The compete calibration plus validation data set consistsof 1700 spectra representing five such measurements for each sample.3.2.3 Spectral Analysis and Chemometric AnalysisWe process spectra collected by Raman measurement using programs written inMatlab to perform background subtraction, wavelet transform, feature selectionand multivariate analysis as described below.3.2.4 Spectrochemical Data Processing and ClassificationAfter mean-centring and normalization, we prepare each spectrum for analysis bydiscrete wavelet transform (DWT). For the purposes of feature selection, we ap-ply TOGA to the transformed spectra. We then develop multivariate, partial leastsquares (PLS) classification models for the prediction of analyte concentrations us-ing DWT and DWT-TOGA databases. We determine the effectiveness of featureselection by comparing residual mean standard errors of prediction (RMSEP) ofmodels based on these two datasets (see below).3.2.5 Discrete Wavelet TransformationDiscrete wavelet transform produces a multiresolution, numerical representationof a spectrum as a sequence of wavelength-space convolutions distributed overdiscrete intervals of position [231, 359]. For the present purposes, we express both49wavelet bandwidth and position by reference to the pixel column number, t, in thespectrum as dispersed on the CCD detector, which relates linearly to the Ramanscattered wavelength, λt .Starting with an orthonormal basis wavelet,ψ j,k(t) = ψ( t2 j− k), (3.1)DWT applies a series of low-pass and high-pass filters to project the spectrum intoa cascading space of paired detail and approximation subbands to form a decom-position,f (t) =∑kc j′,kφ j′,k(t)+j′∑j=1∑kd j,kψ( t2 j− k), (3.2)by which we can represent the spectrum in terms of coefficients, c j′,k and d j,k [143].For present purposes, we use Symlet-5 as a basis function with six levels ofdecomposition. Thus, j runs from 1 to 6, and k extends across the spectrum from1 to 1400. We remove the residual approximation, ∑k c6,kφ6,k(t), which describesthe broad wavelength dependence of the background. We also discard the toptwo detail components, which contain high-frequency noise. DWT thus serves toflatten the baseline and remove noise without distorting the mid-range vibrationalinformation. In this way, we separate informative part of the spectrum into bandsfor multivariate analysis and processing.Template Oriented Genetic Algorithm (TOGA) Feature SelectionFollowing DWT, we apply TOGA to select features at the sub-band level to findspectral information that best correlates with target variance. The multiresonantcharacter of the the DWT decomposition enables an optimization that considersboth feature position (Raman shift) and shape (wavelength-space bandwidth).This process involves three steps. We refer to the calibration data in the matrixof spectra, X, as represented by retained DWT coefficients, d j,k, together withthe corresponding matrix of standardizing property measurements, Y, to developa possibility template. We use this template to form a pool from which to selectinitial genes for TOGA. Having determined the fittest descriptors, we reconstructa DWT representation of the feature-selected spectrum for use in constructing a50multivariate regression model.Possibility TemplateThe possibility template directs the initial selection of variables, consisting of in-dex pairs, j,k judged to be most important. Using Monte Carlo sampling, werepeatedly segregate the spectra contained in X, allocating 75 percent to a trainingset and 25 percent to a validation set. We perform this allocation 300 times, de-veloping 300 different partial least squares (PLS) regression models for a targetedproperty (e.g. concentration of glucose)[94].Each PLS model yields a distinct set of parameters β j,k, the PLS regressioncoefficient for each wavelet, as indexed by j,k [148]. From the families of β j,k,we define the importance of each wavelet coefficient index j,k in terms of a fitnessparameter, s j,k, defined by,s j,k =mean(β j,k)std(β j,k), (3.3)where mean(β j,k) defines the mean value of β j,k and std(β j,k), its standard devia-tion over 300 PLS models. The largest values of s j,k determine a template contain-ing a number of variables (wavelet coefficient indices), p, deemed most important.This possibility template, of a typical dimension, p = 200, forms a predictor spacewith which to orient genetic algorithm processing.Template Oriented Genetic AlgorithmTOGA now begins by choosing ten wavelet coefficient indices, j,k, from the tem-plate of p most important variables. For the choice of eight of these indices, webias our selection to reflect the magnitude of s j,k. We select the remaining two atrandom. Repeating this process 50 times, we form 50 chromosomes. Each chro-mosome defines unique pattern of ten selected wavelet coefficient indices.Applying these to all of the available spectra, we obtain 50 different, feature-selected data matrices, X. For each data such dataset, we use 75 percent of thefeature-selected data to build a PLS calibration model and 25 percent of the datato validate it. We use these calibrations and validations to define a fitness functionfor each chromosome based on an ensemble error criterion determined by the sum51of the RMSEP for cross-validation of the calibration set (RMSEP-CV) and theRMSEP of validation (RMSEP-V).Applying the ensemble error criterion, we select the fittest 30 percent of thechromosomes to pass to the next generation without change. From the remaining35 chromosomes, we select the 18 most fit and perform a two-point crossover togenerate 36 offspring. We discard one at random, and combine the 35 remainingwith the 15 unmodified ones. We mutate five of these at random, drawing fromgenes in the predictor space.Using this modified set of chromosomes, we build a new set of 50 PLS models,generating a new set of fitness functions. We repeat this cycle for 10 generations,saving the final top chromosome, as determined by the ensemble error criterion.Following ten iterations of this entire process we examine the genes in eachcase encountered in the final top chromosomes. Each gene that appears in morethan one top chromosome points to a wavelet component that is retained uponTOGA feature selection.In our experience, by operating in a wavelet transform basis, TOGA serveswell to overcome weaknesses in previous methods for reducing the dimensionalityof spectral datasets defined by a very large number of discrete intensities. Forexample, TOGA yields a continuous spectrum of intensity versus wavelength, asopposed to conventionally selected features, which often appear as non-consecutivespectral elements, making it difficult to extract chemical information from targetedband positions.3.2.6 Reconstructing a Feature-Selected SpectrumTOGA exploits the functional character of the spectrum to identify the most impor-tant features with reference to a basis of wavelet coefficients. Having establishedthis map, we can use it to reconstruct and visualize a reduced, feature-selectedspectrum from any DWT factored experimental spectrum. We simply combine theSymlet-5 functions with coefficient indexes identified by the TOGA selected val-ues of j,k, using the wavelet amplitudes determined by the particular values of d j,kthat describe the spectrum of interest [148].523.2.7 Multivariate CalibrationWe use PLS regression to construct a prediction model by which to examine theeffectiveness with which a selected feature set determines the concentration of spe-cific sugar in a mixture of different sugars.To build a linear regression model, PLS projects the predicted and observablevariables in a new space that maximizes covariance between directions in the mul-tidimensional spaces of X and Y [348]. This method enables modelling success insituations for which the number of observed variables determining X greatly ex-ceeds the number of standardizing property measurements contained in Y. It alsotolerates multicollinearity in X, as present when basing models on largely parallelspectra [114].Various criteria serve to describe prediction success. We refer to, r2, the cor-relation coefficient in the linear regression of predicted versus standard values tojudge the precision of calibration. The root mean square error of prediction (RM-SEP)RMSEP =(m∑i=1(yˆi−yi)2/m)1/2(3.4)describes how well a model, applied to an independent standard data set X, predictsthe corresponding values, Y. Here, yˆi and yi refer to predicted and measured prop-erties of the sample i and m is the number of samples [40]. Examining the standarddeviation about the mean RMSEP value derived from the independent validationof a large number of different calibration reflects the reliability with which a testedpretreatment method improves the accuracy of prediction.For the present work, we assess the calibration error using Monte Carlo crossvalidation (MCCV). This repeated random sub-sampling validation procedure di-vides the large datasets obtained for standard samples prepared over a range ofknown analyte concentration to form a calibration set (75%) and a validation set(25%) [352]. A PLS regression model for the calibration set yields r2, and predictsanalyte concentrations for the validation set, which determines the RMSEP for thatmodel. We repeat this process 50 times to find converged average values for r2 andRMSEP, which we report as a percent coefficient of variation, normalized to theaverage analyte concentration.53Increasing the number of latent variables or PLS factors usually increases thecorrelation between known and predicted values. However, the use of too manyfactors causes over-fitting, which degrades the generality of a model. The numberof samples used to build a model, together with the number of determinant featuresin the spectrum combine to define an optimum number of latent variables.3.3 Results and Discussion3.3.1 Feature-Selected Spectra of Pure Solutions and RandomMixturesFigure 3.2 shows mean-centred, normalized Raman spectra of pure aqueous glu-cose solutions and glucose in aqueous mixtures with random additions of mannose,galactose, and arabinose. This figure also shows corresponding discrete wavelettransforms.We have chosen these monosaccharides to span a range of structural similarity.Mannose and galactose, both hexose sugars, represent C-2 and C-4 epimers ofglucose. Arabinose has a distinctive structure with five carbons. These structuralsimilarities give rise to strongly overlapped spectra.Comparing the spectra of mixtures in Figure 3.2 with the pure monosaccharidesin Figure 3.1, we find no feature that stands out as a distinct univariate marker forany one monosaccharide, with the possible exception of the peak at pixel 600 forglucose. We have constructed a prediction model for the mixed system of foursugars based on this one intensity. This univariate model produces no better errorof prediction than a multivariate calibration using the raw spectra represented inFigure 3.2. From this we can conclude that the feature at pixel 600 incompletelyreflects the information important for determining glucose in mixture of other sug-ars.Figure 3.3 focuses on the glucose system and features isolated by the applica-tion of a TOGA template keyed to the variance in the concentration of this sugar.A collection of calibration spectra of pure aqueous glucose solutions (Figure 3.3a)produces a distinctive pattern of TOGA-selected wavelets shown as Figure 3.3c.The same prescription applied to a calibration set of glucose samples containing540 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel NumberStudent Version of MATLAB0 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel numberStudent Version of MATLABa) b)c) d)0 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel NumberPixel u berFigure 3.2: Mean-centred, normalized Raman spectra of a) 20 aqueous solu-tions of pure glucose over a range of concentrations from 0.04 to 0.10 gml−1. b) all 240 aqueous solutions of glucose galactose, mannose andarabinose mixed in random amounts over a range of concentrations from0.04 to 0.10 g ml−1. c) spectra reconstructed from the DWT-TOGA fea-ture selection of a). d) spectra reconstructed from the DWT-TOGA fea-ture selection of the 60 sample calibration set for glucose with randomadditions of galactose, mannose and arabinose (Figure 3.3b). Spectra,plotted by pixel number represent a Raman shift interval from 320 to2250 cm−1.random amounts of added galactose, mannose and arabinose (Figure 3.3b) isolatesthe set of selected features, as shown by the wavelet transform in Figure 3.3d.Thus, we see that the TOGA spectrum isolated from spectra of mixed samplesby feature selection keyed to glucose, matches remarkably with the TOGA am-plitudes obtained from the set of pure samples prepared over the same range of550 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel Numbera) b)c) d)0 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel NumberStudent Version of MATLAB0 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel NumberFigure 3.3: Mean-centred, normalized Raman spectra of a) 20 pure glucosesolutions in a range of concentrations from 0.04 to 0.10 g ml−1. b) 60glucose solutions in a range of concentrations from 0.04 to 0.10 g ml−1and random added amounts of galactose, mannose and arabinose. c)and d) spectra reconstructed from DWT-TOGA feature selection of a)and b).glucose concentrations. This serves to prove the effectiveness of TOGA for iso-lating the spectrum of a pure substance from the signal produced by a mixture,even one with strongly overlapping spectral features. Note that TOGA emphasizesstructure in a window about the resonance located at pixel 600. This demonstratesthat TOGA responds to evident features when they correlate uniquely with a prop-erty of interest.563.3.2 Effect of Feature Selection on the Prediction Accuracy ofMultivariate Regression ModelsWe gauge the effect of TOGA feature selection on the accuracy with which we canclassify a spectrochemical dataset by comparing PLS models for monosaccharideconcentration based upon the discrete wavelet transform of entire spectra versusTOGA feature-selected DWT spectra. Thus, with reference to the present example,we seek to build calibration models for the concentration of the target monosaccha-ride, glucose, on the basis of its covariance with the full Raman spectral response(Figure 3.2d), and its covariance with the TOGA feature-selected spectral response(Figure 3.3d).−0.02 0 0.02 0.04 0.06 0.08 0.1 0.12−0.0200. concentration (gr mL−1)Predicted concentration (gr mL−1 )  Student Version of MATLAB−0.02 0 0.02 0.04 0.06 0.08 0.1−0.0200. concentration (gr mL−1)Predicted concentration (gr mL−1 )  Student Version of MATLABa) b)Figure 3.4: PLS model using a) raw Raman spectra, b) TOGA reconstructedspectraFigure 3.4a shows the PLS regression of the full Raman spectral responseon glucose concentration for mixed monosaccharide samples in the calibrationdataset. This model exhibits significant non-linearity, with poor sensitivity at bothlow and high concentrations of glucose. Monte Carlo cross validation yields anaverage regression coefficient of 0.74 and a relative RMSEP of 5.4%.The models developed using TOGA feature-selected datasets for calibrationand validation perform significantly better. In Figure 3.4b, we see that predictedglucose concentrations relate to their measured values with a high degree of linear-ity and a much smaller error of validation. As detailed in Table 3.1, we find thatr2 rises to a value near one and the RMSEP falls to 0.74%. These results confirm57that TOGA serves well to filter uncorrelated features without removing importantinformation.We have repeated this same procedure for the other monosaccharides. In eachcase, examining spectra of calibration sets containing randomly varying amounts ofother sugars, we can observe seeming systematic variations in spectral shape withtarget analyte concentration. The TOGA reconstructed spectra we obtain for eachmonosaccharide appear to adapt well to these variations, while excluding parts ofspectrum with large uncorrelated variance and noise. From these qualitative obser-vations, we gain some confidence that TOGA diminishes the weight of uninforma-tive structure and maximizes the weight of features most strongly correlated withtarget variance.Thus, just as we noted above that glucose variation appears to affect the inten-sity of a feature in the DWT spectrum in the region of pixel 600, we can recognizea pair of features for galactose between pixel numbers 300 and 390. Arabinose andmannose both exhibit a relatively intense feature in the region of pixel 200, whichis absent for the other sugars.Figure 3.5 shows that TOGA removes the water signal from all reconstructedspectra, and chooses judiciously among specific combinations of features that ap-pear to signal for each monosaccahride. For example, the calibration spectra foreach of our mixed sugar systems show activity in the region of low-frequency Ra-man shifts from pixels 0 to 150, but TOGA selects this region as important onlyfor galactose.Table 3.1: Regression quality and prediction accuracy of multivariate modelsfor target sugar in mixtures with varying backgrounds of other sugarsbased on complete and feature-selected DWT Raman spectraSugar DWT DWT-TOGAPLS factor r2 RMSEP PLS factor r2 RMSEPGlucose 5 0.74 0.0135 5 1 0.0018Mannose 5 0.85 0.0079 5 0.98 0.0012Galactose 5 0.96 0.0088 7 1 0.0027Arabinose 10 0.92 0.0031 10 1 0.00098580 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel Number0 200 400 600 800 1000 1200Pixel Numbera) b)c) d)Figure 3.5: Toga reconstructed spectrum for a) Glucose, b) Galactose, c)Mannose, d) ArabinoseTable 3.1 gives correlation coefficient of calibration and the root mean squareerror of prediction (RMSEP), for the multivariate prediction model developed foreach of the tested monosaccharides, before and after TOGA feature selection.Prediction performance improved dramatically in all cases. This occurs becauseTOGA decreases the number of variables in the model and selects features thatcorrelate most with the response. Accuracy increases because target- directed fea-ture selection extracts sufficient important information to construct reduced spectrawith few uncorrelated features.in an important confirmation of this methodology, we demonstrate its effective-ness for five independent samples prepared to contain known amounts of differentmonosaccharides, selected randomly to fall in the linear range. Applying TOGA,we reconstruct spectra individually targeted to each target. Building PLS prediction59models to analyze the reconstructed spectra for each monosaccharide we obtain theregression coefficients and root mean square errors of prediction presented in Table3.2. The results show that TOGA models developed independently can be appliedwith high precision to determine the concentration of each component in a newrandom mixture.Previous methodologies for feature selection by projecting spectra into a lower-dimensional space have faced problems, which we believe TOGA overcomes. Onesuch problems arises because the selected features do not adapt well to the naturalshape of the spectra, diminishing the effectiveness with which multivariate analysiscan exploit the information isolated by feature selection. Operating in a space ofwavelet indices, the data-driven genetic algorithm approach presented here isolatesinformation well-adapted to the natural shape of the Raman signal.Table 3.2: Residual mean error of prediction observed in applying individualTOGA models keyed to glucose, galactose, mannose and arabinose tofive aqueous validation mixtures of these sugars containing the randomconcentrations listed below.Sample No. Concentration g ml−1Glucose Mannose Galactose Arabinose1 0.4 2.8 2.6 0.52 1.2 5 3.2 2.43 3.4 2 2 0.94 5 1.6 1.5 35 10 3.9 0.8 1.1Sample No. RMSEPGlucose Mannose Galactose Arabinose1 0.0012 0.0016 0.0028 0.00122 0.0025 0.0011 0.0031 0.00143 0.0022 0.0016 0.0027 0.00124 0.0020 0.0021 0.0030 0.000975 0.0019 0.0009 0.0023 0.00091603.4 ConclusionThis study has introduced a data-driven genetic algorithm feature-selection ap-proach that we term TOGA as a means to identify and quantify the Raman spectraof systems of analytes with overlapping spectral features and uncorrelated vari-ance. The application of TOGA in the space of discrete wavelet transform indicesimproves the prediction ability of the model by adapting TOGA feature selectionto the shape of the shape the Raman resonances. Improvement in the performanceof prediction models after applying TOGA proves that this method extracts recon-structed spectra that offer important and sufficient information for classification.TOGA reconstructed spectra invoke signal features noticeable in the variation ofmixtures with determinate variance, suggesting a utility for associating variationsin classified properties with chemical substances.61Chapter 4Multivariate classification of pulpNIR spectra for end-productproperties using discrete wavelettransform with orthogonal signalcorrectionNatural material variations uncorrelated with physical properties of fibre networkshinder the development of robust calibration models by which to predict paperproperties from on-line near-infrared (NIR) spectra of production pulps. Such asimple process gauge of product quality would offer attractive advantages for opti-mized manufacturing. The present work explores the effectiveness of data process-ing strategies designed to remove uncorrelated variance from calibration modelslinking NIR spectra with standard measures of paper quality, including tensile,tear, burst strength, wet and dry zero span length, freeness, absorption and scatter-ing coefficients. Post processing of spectra by discrete wavelet transform (DWT)is shown to suppress baseline and high-frequency noise, and orthogonal signal cor-rection (OSC) substantially improves prediction accuracy by reducing the ampli-tude of uncorrelated (orthogonal) variations. We find that combined pretreatment62by DWT and OSC yields a spectral data set that exhibits the best prediction accu-racy.4.1 IntroductionTo secure market superiority, a pulp manufacturer must control the quality of ma-terials in process to maximize product sheet strength. While some pulp propertiescan be tested at line, conventional determinations that best predict the physicalproperties of an end-use product require exacting measurements in a controlledlaboratory environment. Such steps add cost and introduce delay that can give riseto process variability.A need therefore exists for on-line methods to gauge pulp stream composi-tion and morphology that can predict the structural and physical properties of theend-use paper it forms. To this end, the industry has sought to develop spec-troscopic probes.[315] Chief among such methodologies is near-infrared (NIR)spectroscopy.[34, 193, 315]NIR diffuse reflectance absorption spectroscopy has been applied with successfor the analysis of lignocellulosic materials, particularly chipped wood feedstocks,[274,356] for physical properties, such as moisture content,[172, 310, 317] density,[103,268, 327] surface roughness,[104, 318] as well as levels of lignin,[241, 242, 274,319] cellulose,[124, 350] hemicellulose[274, 316] and other extractives.[241, 316]Regression models based on NIR spectra of laboratory test sheets have shownpromise as a means to predict conventionally measured pulp and paper properties,such as freeness, stretch and tensile strength.[88, 89]Two factors limit the full-scale application of NIR spectroscopy as an on-lineprocess-control tool. Broad overtone absorption bands, owing to every substancein the sample, overlap to form a spectrum in which signature features are hard todiscern. Variations in pulp composition unrelated to a target property modulate thespectrum, and this tends to mask determinate spectral variations.The present work reports progress in an effort to overcome these limitations.Taking a large NIR data set of production pulps, standardized by conventional mea-surements of product sheet properties, we show that multi-resolution decomposi-tion by discrete wavelet transform (DWT) effectively recasts spectra to facilitate63the isolation of determinant variance. Preprocessing calibration spectra by orthog-onal signal correction (OSC), we minimize uncorrelated variance, which enablesthe development of multivariate classification models that predict paper sheet prop-erties from pulp NIR spectra with greatly improved accuracy.4.2 Materials and Methods4.2.1 Physical, Mechanical and Optical Properties of Pulp SamplesSeventy five test sheet samples were produced from production pulps in the Burn-aby laboratories of Canfor Innovation. Standardizing information available forthese unrefined pulps include tensile strength, freeness, burst index, dry zero span,wet zero span, tear index, specific refining energy to reach a freeness of 600 (SRE600),absorption and scattering. Table 4.1 gives the range of each parameter as well asits measured uncertainty.Table 4.1: The experimental range of measurement and observed repro-ducibility of different the physical and optical properties independentlydetermined for the samples that served as calibration and validation stan-dards for this study.parameters Minimum Maximum Mean Repeatability ReproducibilityTensile (km) 2.5 4.9 3.8 5 % 10 %Freeness (ml) 649.5 700.5 684.3 25.3 % 32.4 %Burst (kPa m2 g−1) 1.4 3.4 2.3 22 % 28 %Dry Zero Span (km) 14.44 16.8 15.7 5 % 10 %Wet Zero Span (km) 12.4 15.8 14.6 5 % 10 %Tear (mN m2 g−1) 19.9 30.3 26.1 4.2 % 12.5 %SRE600 23.5 82.7 53.4 — —Absorption (m2 g−1) 0.1669 0.2278 0.1901 — —Scattering (m2 g−1) 32.6 39.8 35.6 — —644.2.2 Collection of NIR SpectraWe collected NIR spectra using a Nicolet 6700 FT-IR spectrometer (ThermoScien-tific) equipped with an NIR integrating sphere module and a 5 cm diameter samplecup spinner. Operated in the diffuse reflectance mode, this instrument illuminatessamples with broadband near-infrared radiation from a tungsten halogen lamp, andcollects interferograms using an InGaAs detector. Fourier transforms span a spec-tral range from 4000 cm−1 to 9900 cm−1 at a resolution of 2 cm−1. Each acquisi-tion represents the sum of 64 scans of 10 second exposure.We cut five circular samples sized to fit the sample cup from different positionsin each pulp sheet, scanning the top-side of each. Five spectra from each samplewere averaged and used in subsequent analyses.4.2.3 Multivariate Analysis4.2.4 Data PretreatmentNIR spectral data can benefit to a great degree from preprocessing to compensatethe effects of varying baseline, remove high-frequency noise and amplify featuresthat appear only as inflections in the raw spectrum. The present work tests severaltarget-independent preprocessing methods, including derivatization, standard nor-mal variant (SNV) correction, multiplicative scatter correction (MSC) and discretewavelet transform (DWT). Each of these approaches refines the data set withoutreference to standardizing classification information.We compare prediction errors following pretreatment with these methods withresults obtained following orthogonal signal correction (OSC) alone, and OSC incombination with DWT. OSC is a target-directed preprocessing method that makesreference to a multivariate classification model to reduce the weight of irrelevantspectral information.4.2.5 First-Derivative TransformationNIR absorption spectra convey sample composition information in the form ofoverlapping vibrational overtone features that vary smoothly on a scale of tens ofwavenumbers. First derivative transformation provides a ready means of identify-65ing such features, even when their appearance in the primary spectrum is subtle.[102,115] The first derivative also serves to highlight the degree to which high-frequencyproperties of the spectrum, unrelated to variations in the sample set, might affectits classification.[7, 115]4.2.6 Standard Normal Variant CorrectionStandard Normal Variate (SNV) transformation operates by row on spectra, xi inthe data matrix X, subtracting the individual mean (zeroth-order detrending) andscaling each spectrum by its standard deviation.[100]xˆi =xi− x¯isi(4.1)When much of the amplitude fluctuation in a data set arises from noise, thispretreatment transformation can improve a model by reducing the amplitudes ofits noisiest component spectra. However, if overall signal amplitudes of sam-ple spectra vary significantly SNV can degrade calibration by introducing non-linearities.[29]4.2.7 Multiplicative Scatter CorrectionMultiplicative scatter correction (MSC) removes uncorrelated background frommeasured spectra, X, arising from multiplicative factors, such as path-length vari-ations, as well as offsets, owing, for example to stray light. The procedure usesan averaged spectrum, x¯j, formed by the mean of a selected calibration subset, andfinds parameters, ai, bi and ei, fitting xi to x¯j as closely as possible by least squares:xi = ai+bix¯j+ ei (4.2)where ei represents the residual spectrum, containing the chemical information inxi. ai defines the intercept, and bi the slope, that yield the corrected spectrum,xi,MSC:xi,MSC =xi−aibi= x¯j+eibi(4.3)66This procedure presents a risk in that ai and bi might correlate with a target prop-erty, in which case MSC would remove chemical information from the data set.[13,100, 194, 195]4.2.8 Discrete Wavelet TransformDiscrete wavelet transform produces a multi-scale representation of a digitizedsignal by using a sequence of high- and low-pass cutoff filters to sort the sig-nal in terms of the frequency with which it varies in the wavelength space ofthe spectrum. Filtering divides this information to resolve the signal into a setof subbands.[49, 80, 329]Frequencies of importance, corresponding to peak widths in the wavelength-or λ -space of the original spectrum, appear with large amplitude in the DWT de-composition without loss of λ -space position information. Subsampling the resultremoves the unimportant information, including the slowly varying backgroundand high-frequency noise. Errors of prediction can serve as a guide to choose awavelet basis and constrain frequencies to best preserve classification informationwhile removing uncorrelated variance.[49]4.2.9 Orthogonal Signal CorrectionEven though multivariate classification models such as Partial Least Squares (PLS)regression serve to extract feature information from complex data sets, uncon-trolled variations, irrelevant to the properties of interest can add substantially to thecomputational effort, reducing the accuracy and robustness of prediction. Some-times, uncorrelated variation appears in obvious dimensions, such as low-frequencybaseline oscillations, or high-frequency noise. In other cases, irrelevant variationsintrinsic to the sample or the instrument occur at the bandwidth of the determinantvariation.In the latter event, preprocessing methods focused on the minimization of un-informative variation at all frequencies can serve to improve the accuracy of a mul-tivariate calibration. Orthogonal Signal Correction (OSC) provides one approachto succeed in this respect.[347]OSC finds features that affect the total variation in a spectral matrix, but occur67in dimensions that extend in directions orthogonal to a target variance, and thenremoves them. Omitting these features reduces the complexity of the model andconsequently improves the linearity of the relation between a target variance andan input data set.Our implementation of OSC applies Principal Component Analysis (PCA) to a75 percent subset of the spectral data in X in order to obtain an initial score vector,t. We orthogonalize t with respect to the corresponding target property matrix,Y. This orthogonalized t∗ then serves as the first target in a subsequent cycle ofsupervised refinement.Using PLS regression, we find weights, w, such that the vector of PLS scorest = Xw conforms optimally with t∗. We orthogonalize this t with Y, and, withthis as a target, return to PLS to obtain a new t. Upon convergence, the vector ofscores, t, and loadings, P, from this process describes the information in X thatis orthogonal to Y. Applying the transpose of P yields a feature-selected dataset,XOSC by:XOSC = X− tP′ (4.4)XOSC then serves as spectral data input for a new cycle of OSC, generating neworthogonalized component, tnew and Pnew, with which to feature-select Xnew. Totreat the NIR spectra of the present set of pulp samples, we have found it best toimplement OSC as it is most conventionally applied, stopping with two such scoreand loading components.[301]Multivariate CalibrationWe use the method of Partial Least Squares (PLS) regression to compare efficacyof these preprocessing methods for classifying physical properties of productionpulps on the basis of NIR spectra. To build a linear regression model, PLS findsthe covariance between a set of observed variables (the matrix of NIR spectra, X)and predicted variables (the matrix of values of a selected property, Y). It rotatesX into a coordinate direction that best represents the variance in Y.[99, 348]This method enables modelling success in situations for which the numberof observed variables determining X greatly exceeds the number of standardizing68property measurements contained in Y. It also tolerates multicollinearity in X, aspresent when basing models on largely parallel spectra.PLS reduces the dimension of data into latent structures in the X block ac-cording to a selection of principal components. The algorithm calculates scoresof blocks to maximize the co-variance between X and Y. The weight vector cal-culated for each PLS component gives the maximum covariance between the twoblocks.[114]Various criteria serve to describe prediction success. We refer to the correlationcoefficient in the linear regression of predicted versus standard values to judge theprecision of calibration. The root mean square error of prediction (RMSEP)RMSEP =(m∑i=1(yˆi−yi)2/m)1/2(4.5)describes how well a model applied to an independent standard data set X predictsthe corresponding values, Y. Here, yˆi and yi are predicted and measured prop-erties of the sample i and m is the number of samples. Examining the standarddeviation about the mean RMSEP value derived from the independent validationof a large number of different calibration reflects the reliability with which a testedpretreatment method improves the accuracy of prediction.Increasing the number of latent variables or PLS factors usually increases thecorrelation between known and predicted values. However, the use of too manyfactors causes over-fitting, which degrades the generality of a model. The numberof samples used to build a model, together with the number of determinant featuresin the spectrum combine to define an optimum number of latent variables.[99, 132]For the purposes of evaluating the pre-processing strategies described above,we constructed and tested sets of classification models using PLS. We began bybuilding nine Y matrices, detailing the known pulp physical characteristics, tensileindex, freeness, burst index, dry zero span, wet zero span, tear index, SRE600,absorption and scattering. We then applied PLS to find latent components in Xmatrices of preprocessed spectra characterizing the spectrum-property covariance.For each property and preprocessing strategy tested, we used 75 percent of theavailable spectra for calibration and the remaining excluded 25 percent as a test set69to validate the model.PLS requires one to choose an optimal number factors minimizing predictionerror without over fitting the data. Interestingly, we generally found that the op-timal number of PLS factors declined after pre-processing. This reflects the factthat preprocessed spectra had less complexity, necessitating fewer PLS factors tooptimally capture the covariance.4.3 ResultsThe following sections draw upon a data set formed by the near infrared spec-tra of 75 pulp test sheets, comparing the effects of pre-treatment by first deriva-tive and standard normal variant (SNV) transformation, multiplicative scatter cor-rection (MSC), discrete wavelet transform (DWT), orthogonal signal correction(OSC), and the combination of DWT with OSC. We explore the comparative util-ity of these pretreated data sets as multivariate predictors of various mechanicaland physical properties of sheet paper, including tensile, dry and wet zero-span,tear and burst index, freeness, SRE600 and light scattering.We have constructed these models and tested prediction accuracy by applyingthe method of partial least squares to independent pretreated data subsets. Here,we compare RMSEP values obtained by use of preprocessed spectral data withprediction errors found for PLS models based on the original spectra without pre-treatment.4.3.1 NIR Spectra with Target Independent PretreatmentFigure 4.1 plots representative, mean-centred and normalized NIR spectra of all75 pulp samples, together with this dataset pre-processed by non-targeted first-derivative transformation, SNV and MSC. In the untreated data, we see that ab-sorption bands associated with many vibrational overtone transitions overlap toform smooth, reproducible and, on this scale, relatively undifferentiated spectra.Less apparent in the raw NIR spectra is the presence of a high-frequency noisecomponent. This contribution to the experimental signal plainly appears upon first-derivative transformation. No NIR spectral intensity varies on this scale of absorp-tion frequency. Therefore, this evident variation in the dataset cannot correlate with704000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (a)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (b)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (c)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (d)Figure 4.1: (a) Representative NIR spectrum in cm−1 for each of the 75 pulpsamples in the present dataset (c.f. Table 4.1) (b) Spectral data subjectedto first-derivative transformation. Spectral data subjected to nnnnnnn(c)standard normal variant (SNV) transformation and (d) multiplicativescatter correction (MSC).the variation in any property of the pulp samples.Standard normal variant (SNV) transformation and multiplicative scatter cor-rection (MSC) have similar effect on the NIR data set. Comparing plots in Figure4.1, we find that both of these pretreatment methods highlight structural featureswhile reducing baseline offsets and high-frequency noise.Discrete wavelet transform enables the selective application of low and high-frequency filters that remove baseline offsets and noise with great effectiveness.Figure 4.2(a) plots the full dataset after a seven-level decomposition in a Symlet-5basis followed by the removal of the three highest frequency components (details),71and the lowest one (approximations).4000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (a)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (b)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (c)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1) (d)Figure 4.2: (a) Spectra following a discrete wavelet transform of the rawspectral data in Figure 4.1 (b) The same spectra following orthogonalsignal correction (OSC), targeted to a prediction of the paper sheet me-chanical property, tensile strength. (c) Raw spectral data pretreated firstby discrete wavelet transform, as above, followed by two-componentorthogonal signal correction targeted to tensile strength. (d) Pretreatedspectral data reversing the order of transformation by OSC and DWT.Here we see an elimination of baseline slope and offset, together with a signifi-cant amplification of reproducible structure over the full range of the spectrum andthe elimination of high-frequency noise.724.3.2 Effects of Target-Directed PretreatmentAll of the data preprocessing methods illustrated above can serve to isolate struc-ture and reduce noise in the NIR spectra of cellulosic pulps. However, these global,non-targeted preprocessing strategies do not distinguish between a spectral featurethat correlates with a target variation in a mechanical or physical property of thepaper made from these pulps and one that does not.Target-directed preprocessing techniques make reference to measured stan-dards in order to select features that enhance the determinate variations in a dataset, while suppressing variations that are uncorrelated or orthogonal to a targetedproperty. We explore the effectiveness of such a feature selection strategy for theNIR classification of pulps by applying the method of orthogonal signal correction(OSC).4.4 Discussion4.4.1 Summary Overview of Pretreatment EffectsFirst-derivative preprocessing, as displayed in Figure 4.1(b), clearly increases thehigh-frequency noise in the spectral range above 7000 cm−1. However, as dis-cussed below, we find its elimination of the baseline improves the prediction mod-els that we can construct for most target parameters. Background and noise bothcontribute to limit the accuracy of a prediction model, and for the present case, sta-bilizing the baseline outweighs the introduction of completely uncorrelated high-frequency noise.SNV and MSC pretreatment (Figure 4.1(c) and (d)) yield spectra with a similarstructure. Neither method enhances resolution or discrimination. They do decreasethe high-frequency noise. But they leave spectra with background oscillations thatdo not correlate with target variance.DWT operates in a more deterministic way to decompose the spectrum intosubcomponents based on the bandwidth of its feature variations. Each such com-ponent represents different domain of information. The conformance of features inFigure 4.2(a) with vibrational overtone line shapes, shows how a choice of band-width scale can substantially reduce baseline variations and high-frequency noise73while highlighting determinate information.Pretreatment by OSC sharpens resolution compared with raw spectra, and servesto accentuate variation relevant to target properties. However, its implementationhere introduces noise, particularly above 9,000 cm−1, and fails to suppress thebackground.Thus, we see that wavelet transform offers an effective means to subtract back-ground and remove noise, while OSC suppresses irrelevant information. Seekingthe discrimination enhancement afforded by OSC, for spectra filtered against back-ground and noise, we have explored combinations of these prepossessing methods.Figure 4.2(c) shows that OSC applied to a spectral data set pretreated first by DWTyields a baseline-stabilized NIR signature with apparently enhanced signal vari-ance, but also some increase in high-frequency noise.Figure 4.2(d) reverses the strategy. Here we remove orthogonal variance toenhance intensity differences that correlate best with the target. Then we de-noiseby applying DWT. This produces a sparser representation, with less high-frequencynoise and better isolated variation within the bandwidth of the molecular NIR spec-tral response.To assess the effectiveness with which preprocessing improves the accuracy ofclassification we have built multivariate models from subsets of pretreated spectra,and run independent validations to estimate residual mean errors of prediction.4.4.2 Prediction of Physical Properties on the Basis of NIRClassification ModelsOur study examines the effectiveness of various methods for preprocessing NIRspectra of pulp on the success of multivariate prediction models for nine phys-ical properties of paper, including tensile, freeness, burst, DZS and WZS, tear,SRE600, absorption, and scattering. We have constructed PLS models and calcu-lated the accuracy of prediction for each property to evaluate the efficiency of eachpreprocessing method.Ultimately, the accuracy that any of these assessment can have is limited byaccuracy of standard measurements used to determine property values of our stan-dards. As indicated in Table 4.1 the measurements that determine the properties ofthe samples used as standards in this study vary in their reproducibility. Bearing74this limitation in mind, we proceed now to compare the relative effect of vari-ous methods of data pretreatment on classification accuracy within each individualproperty.We carry out this comparison by applying different analytical criteria, such asaccuracy, precision, and linearity to evaluate PLS models. With reference to Tables4.3 and 4.2, we evaluate the prediction models for tensile strength using untreatedand first-derivative treated spectra.We see that, using a sufficient number of PLS factors, untreated spectra canserve to yield a linear PLS calibration model. The application of a simple first-derivative pretreatment reduces the number of factors required and improves per-formance at the stage of validation. But, we find that the regression model afterfirst-derivative preprocessing exhibits an intercept that differs more from zero. Ev-idently, the reduction in background with pretreatment decreases scatter, but theincrease in high-frequency noise adds a systematic prediction offset.Using spectral data sets preprocessed by SNV and by MSC. We find a moderateimprovement in tabulated RMSEP values, and PLS regression models with (x,y)intercepts closer to (0,0), suggesting that these pretreatment methods better succeedin removing random variance without adding a systematic offset to the model.Of the remaining two pretreatment methods available to us in this study, DWTprovides a means to filter slowly varying background and highly oscillatory noise,while OSC reduces the dimensionality of the data by suppressing variance in co-ordinates deemed orthogonal to the targeted analysis. Table 4.3 lists the errors ofprediction obtained for paper properties by applying these pretreatments individ-ually. Here we see comparable levels of improvement in RMSEP owing to theremoval of background and noise (DWT) or the amplification of relevant signalcomponents (OSC). OSC, applied alone, yields a slightly better prediction modelfor tensile strength.75Table 4.2: Optimized number of PLS factors, h, and the regression coefficient btained from the model fit to the calibra-tion data-set, following various methods of NIR spectral pretreatment.Parameters Non 1D SNV MSC DWT OSC DWT-OSC OSC-DWTh r2 h r2 h r2 h r2 h r2 h r2 h r2 h r2Tensile 10 0.95 3 0.92 10 0.94 10 0.94 7 0.93 7 0.98 7 1 7 0.93Freeness 10 0.71 3 0.71 7 0.53 7 0.53 7 0.78 7 0.82 7 0.92 7 0.78Burst 10 0.91 3 0.89 10 0.95 10 0.95 7 0.95 7 1 7 1 7 0.91DZS 10 0.84 3 0.79 10 0.83 10 0.83 7 0.86 7 0.92 9 0.96 10 0.93WZS 7 0.74 3 0.57 7 0.68 7 0.68 7 0.9 7 0.91 7 0.98 7 0.86Tear 7 0.59 2 0.7 7 0.7 7 0.7 7 0.79 7 0.97 7 1 7 0.95SRE600 10 0.76 3 0.8 10 0.78 10 0.78 7 0.86 7 0.97 7 0.98 7 0.96Absorption 10 0.83 3 0.87 10 0.86 10 0.86 4 0.88 10 0.96 7 0.99 10 0.96Scattering 10 0.9 3 0.82 7 0.82 7 0.82 7 0.93 9 0.98 7 1 7 0.9576Because pretreatment by DWT and OSC suppress uncorrelated variance in dis-tinctively different ways, it seems reasonable to explore whether there is furtheraccuracy to gain by preprocessing the spectral data both by DWT and OSC. Table4.3 shows the results of such a strategy both for the case in which we first denoisethe data by DWT and then apply OSC, and for the case in which we first suppressuncorrelated variance by OSC and then remove uninformative noise by DWT.We find, by and large, that pretreatment in the sequence OSC-DWT yields cal-ibration models with the lowest RMSEP and smallest intercept for almost all of theproperties tested. Table 4.3 gives average RMSEP value found by the applicationof five distinct PLS regression models (five randomly selected calibration and val-idation datasets) to predict each physical property using NIR spectra preprocessedby each of the methods discussed above. We find that the order, OSC followed byDWT, yields the lower average RMSEP for all properties except freeness.Table 4.3 also reports the standard deviation of the distribution of RMSEP ineach case. This quantity reflects the precision to which we can specify the pre-diction improvement achieved by a given pretreatment strategy. The lower thisstandard deviation, the more consistently the pretreatment strategy operates to im-prove prediction accuracy for a given property, regardless of the conformance thatmight happen to exist between a particular choice of calibration and validationsets. A higher standard deviation indicates an inconsistency in prediction accuracy,reflecting an incomplete suppression of variance in the pretreated data.Table 4.3 shows that OSC yields a relatively broad distribution of RMSEP re-sults, characterized by a larger standard deviation. However, combining OSC withDWT gives the lowest RMSEP of all pretreatment methods for almost all parame-ters. As a targeted preprocessing method, OSC operates best when the validationspectra happen to vary over a data space that is well represented by the calibra-tion spectra (including orthogonal variance). DWT is non-targeted, and preformssimilarly in reducing error regardless of the degree of conformance that happens toexist between the calibration and validation data spaces.Combining DWT with OSC in either order, we see the lowest RMSEP, withthe narrowest standard deviations. Apparently, the application of a wavelet filteracts to reduce random variance, yielding spectra of greater uniformity from whichto draw calibration and validation datasets.77The regression correlation coefficient of a model reflects the stability of predic-tion over the full range of the target property. A model with wide stability yields aregression coefficient near one. Values higher and lower indicate systematic vari-ations for larger or smaller values of a target property. (Table 4.2) shows that thecombination of OSC and DWT improves the regression coefficient significantly forall parameters. We can trace this result to the effectiveness of this pretreatment inremoving systematic uncorrelated variance.4.4.3 Preprocessing as a Means of Isolating Spectroscopic FeaturesNIR spectroscopy presents the inevitable problem of highly overlapped spectro-scopic features. By suppressing uncorrelated variance, all forms of pretreatmentcan serve to accentuate the elements of a spectrum that correlate the variance ofinterest. First-derivative, SNV and MSC preprocessing methods achieve this byreducing noise and baseline variations. DWT flattens baseline, much like first-derivative preprocessing, and removes high-frequency noise. OSC serves to high-light features of determinate variance by subtracting orthogonal information fromthe spectrum.Examining the spectroscopic outcome of pretreatment can provide a visual toolto understand the source of a target variance in terms of its spectroscopic manifesta-tion. The physical properties of a pulp relate to each other by common correlationswith fibre morphology and chemical composition. Feature-selected spectra, builtto conform individually with target properties should therefore exhibit similaritiessignalling these underlying chemical and physical correlations. The spectra ob-tained here, following pretreatment geared to properties known to be correlated,present evidence supporting this idea.78Table 4.3: Average residual mean square errors of prediction following the application of various methods of NIRspectral data pretreatment. The RMSEP averages determined in each case by partial least squares calibrationmodels using ten different randomly selected training and test data sets. Tabulated standard deviations reflect thereproducibility of RMSEP following different preprocessing methods.Parameter PLS 1st-PLS SNV-PLS MSC-PLS DWT-PLS OSC-PLS DWT-OSC-PLS OSC-DWT-PLSTensile (km) 0.2742± 0.2473± 0.2214± 0.2213± 0.1985± 0.1960± 0.1958± 0.1754±(2.5-4.9) 0.0096 0.0061 0.0033 0.0033 0.0019 0.0018 0.0018 0.0013Freeness (ml) 12.04± 11.92± 10.18± 10.17± 9.327± 10.02± 10.03± 10.10±(649.5-700.5) 1.603 0.9726 0.8147 0.8165 0.7857 0.8009 0.7741 0.7437Burst (kPa m2 g−1) 0.2509± 0.2455± 0.2349± 0.2348± 0.2183± 0.2229± 0.2175± 0.2171±(1.4-3.4) 0.0067 0.0048 0.0022 0.0024 0.0022 0.0027 0.0007 0.0016Dry Zero Span (km) 0.4556± 0.4344± 0.4109± 0.4109± 0.4081± 0.4169± 0.4062± 0.4025±(14.44-16.8) 0.0020 0.0023 0.0007 0.0007 0.0009 0.0015 0.0006 0.0008Wet Zero Span (km) 0.5881± 0.5677± 0.5266± 0.5264± 0.5010± 0.5145± 0.5023± 0.5018±(12.4-15.8) 0.0052 0.0045 0.0030 0.0029 0.0022 0.0032 0.0024 0.0022Tear (mN m2 g−1) 1.8163± 1.8001± 1.7421± 1.7418± 1.6994± 1.7099± 1.6898± 1.7061±(19.9-30.3) 0.0046 0.0043 0.0075 0.0079 0.0086 0.0097 0.0038 0.0059SRE600 12.7136± 12.6064± 12.1212± 12.1215± 12.3330± 12.3978± 12.2578± 12.2095±(23.5-82.7) 0.0705 0.0661 0.0439 0.0440 0.0511 0.0713 0.0563 0.0471Absorption (m2 g−1) 0.0094± 0.0093± 0.0090± 0.0090± 0.0085± 0.0087± 0.0082± 0.0084±0.1669-0.2278) 0.0007 0.0007 0.0007 0.0007 0.0005 0.0005 0.0006 0.0005Scatering (m2 g−1) 1.0116± 1.2047± 0.8780± 0.8777± 0.8101± 0.9006± 0.8093± 0.8033±(32.6-39.8) 0.0052 0.0052 0.0054 0.0053 0.0036 0.0050 0.0029 0.005579For example, Figures 4.3 to 4.5 compare NIR spectra after OSC-DTW pre-treatment tied to Burst and Tensile Strength, Wet Zero-Span and Dry Zero-Span,and SRE600 and Tear. In the universe of standard samples for this study, Burstand Tensile Strength, which range over a factor of two, correlate linearly, with aleast-squares correlation coefficient of 0.964. We see this reflected in the feature-selected spectra obtained following OSC-DWT, which exhibit a large variation inthe region of from 4300 to 5800 cm−1 with distinct isosbestic points at 4700 and4900 cm−1.4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)Figure 4.3: Pulp NIR spectra after processing by combination of orthogonalsignal correction with two components and discrete wavelet transformwith selected wavelet range of (3-7) with reference to (left) Burst, and(right) Tensile Strength.Wet Zero-Span and Dry Zero-Span show a similar degree of correlation in ourstandards, but far less variation. OSC-DWT preprocessing tied to either of theseproperties yields a similar set of resonances in the region of 5000 cm−1, whichare very different from the sub-spectra characteristic for Tensile and Burst. Thesefeatures span the WZS-DZS range of our samples with a much smaller amplitudevariation.Finally, SRE600 refers to a pulp property that determines the amount of re-fining energy necessary to reach a Freeness of 600. This property relates to fibrelength much like that characteristic of the fibre network measured by Tear. In oursamples, we find that SRE600 does correlate well with Tear, and spectrally we see804000 5000 6000 7000 8000 9000 10000Frequency (cm−1)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)Figure 4.4: Pulp NIR spectra after processing by combination of orthogonalsignal correction with two components and discrete wavelet transformwith selected wavelet range of (3-7) with reference to (left) Wet ZeroSpan, and (right) Dry zero span.that OSC-DWT selects for a very similar set of features.4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)Figure 4.5: Pulp NIR spectra after processing by combination of orthogonalsignal correction with two components and discrete wavelet transformwith selected wavelet range of (3-7) with reference to (left) SRE-600,and (right) Tear.Among all the pulp properties investigated in this study, Scattering is unique inthat it represents an optical property of the fibre network, which correlates poorly81with indexes derived from measures relating to fibre network strength. As shown inFigure 4.6, amplitude of the features in the NIR spectra of pulp selected to isolateScattering shifts decidedly to the red compared to strength-related elements of thespectrum, and variance concentrates in a resonance at 5100 cm−1.As also shown in Figure 4.6, we find that OSC-DWT selects a unique set offeatures to describe Freeness as well, despite the expectation that, in its morpho-logical and chemical basis, freeness should correlate with other measures relatedto fibre length such as SRE600, Tensile Strength, etc. However, we see from Table4.3 that Freeness represents the only parameter in this study for which predictionfrom the NIR spectrum was not improved by OSC-DWT preprocessing. We con-clude from this that the OSC algorithm, applied with reference to Freeness, failsto capture the most correlated features. For this reason, we must not regard theOSC-DWT selected features as a trustable source for spectroscopic interpretationfor Freeness.4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)Figure 4.6: Pulp NIR spectra after processing by combination of orthogonalsignal correction with two components and discrete wavelet transformwith selected wavelet range of (3-7) with reference to (left) Scattering,and (right) Freeness.4.5 ConclusionWe have explored the effectiveness with which a range of preprocessing methodsimproves the multivariate prediction of paper physical and morphological proper-82ties from the NIR spectra of pulp fibres. Simple data treatments, including first-derivative, SNV and MSC preprocessing do reduce prediction errors. However,DWT, which removes high-frequency noise and low-frequency baseline variations,and OSC, which suppresses variance orthogonal to a specified target property, im-prove performance to a better degree. Interestingly, DWT, which preprocesseswithout reference to a target property, yields regression coefficients that fall shorterof one, indicating the presence of a systematic residual uncorrelated variance.Combining the pretreatment methods, OSC and DWT yields the smallest errors.Using Tensile Strength as a benchmark, the sequence, OSC then DWT producesa more robust prediction, featuring a lower RMSEP with a narrower distribution.Either order improves RMSEP and standard deviation of prediction error to sim-ilar degrees for other properties, For all cases the application of DWT followedby OSC gives a regression coefficient closer to one, suggesting that this approachbest avoids the biasing of noise suppression with respect to magnitude of the targetproperty.NIR spectra corrected by OSC-DWT reveal interpretable features with respectto selected properties. Correlations in such spectra provide a means to logicallyconnect certain properties with assigned features. These results underline the effec-tiveness with which OSC-DWT extracts essential information from spectra subjectto uncorrelated variations. Thus, we can conclude that the application of OSC-DWT as a pre-treatment can improve the utility of a NIR spectroscopy as a methodfor predicting the end-point properties of pulp and paper.83Chapter 5TOGA Feature Selection and thePrediction ofPhysical-Mechanical Propertiesof Paper from the Raman Spectraof Unrefined PulpWe demonstrate that Raman spectra of unrefined pulp sample can accurately pre-dict standard properties of paper made from that pulp at any arbitrary refining en-ergy when modeled to a reduced space selected by a sophisticated chemometricfeature selection technique referred to as a Template Oriented Genetic Algorithm(TOGA). We combine discrete wavelet transform (DWT) with TOGA feature se-lection and partial least squares (PLS) multivariate classification to extract maxi-mum chemical and physical information from raw Raman spectra to correlate Ra-man features with five physical properties of pulp at five different refining energies.We remove uncorrelated features to develop robust calibration models by which torepresent paper properties from on-line Raman spectra of production pulp. UsingTOGA reconstructed Raman spectra more than doubles the prediction accuracyfor certain properties. We also investigate the effects of chemical bonds and the84structure of cellulose fibre on the mechanical properties of refined kraft pulp us-ing TOGA selected Raman signals. Our results indicate that the individual fibrestrength rather than inter-fibre bonding serves as the defining factor for specificrefining energy (SRE), freeness, wet and dry zero-span breaking length while bothsingle-fibre properties and chemical bonding affect sheet physical and mechanicalproperties such as tensile, burst and tear strength. Pulp changes associated with di-minished freeness at increased refining energy reflect increased fibre-fibre bonding,which has a major effect on mechanical properties of pulp. The Raman spectrumof a pulp conveys information both about its chemical composition and fibre struc-ture, providing an effective means to predict the quality of paper sheet products. Inparticular, the facility demonstrated here to gauge the product properties at any ar-bitrary refining energy based on the Raman spectrum of an unrefined pulp can savetime and cost by determining in advance how much beating is required to producea desired set of paper product specifications.5.1 IntroductionA roll of manufactured paper sheet attains a value and fitness for purpose that isdetermined by a suite of physical and mechanical properties. These propertiesvary with characteristics of the feedstock and parameters of production process.To gauge the suitability of a paper product with respect to target specifications,manufacturers rely for the most part on measured properties of the finished goods.Thus, a bag paper must meet or exceed a particular bursting strength specification.A printing paper must surpass a certain tear resistance threshold. All papers mustattain appropriate degrees of tensile strength, particularly for applications that de-pend on stress-strain performance.Directly measured information about physical and mechanical properties of apaper sheet determines the acceptability of a production run. But, off-line labor-intensive laboratory measurements offer no means to alter process conditions toachieve or sustain a product specification.Paper consists of a complex fibre network. The properties of the fibres and thearrangement they form combine to determine the properties of paper. Automatedsensors can provide online process information on the pulp stream (concentration,85freeness, brightness, pH and kappa number) and the fibres it contains (image dataon length, thickness, curl, kink and coarseness). Calibration models have begun toconnect measures such as these to process parameters and product attributes [149].However, the chemical composition of the fibres themselves and their morphologyinterplay to determine the physical properties of a paper sheet. Control in a spacethat encompasses the complete range of variance requires means of sensing bothfibre chemical and physical responses to production conditions [6, 89, 308].5.1.1 Chemical Composition of PulpPapermaking begins with a highly heterogeneous feedstock of biological origin.Pulping and refining extract polysaccharide fibrils composed to varying degreesof cellulose and hemicellulose. Cellulose has the simple linear structure of D-glucose, linked in a repeating sequence of β (1,4) glycosidic bonds [159]. Thehemicellulose component differs in its profusion and composition depending uponthe plant source.Hemicelluloses link a variety different monosaccharides, including glucose,xylose, mannose, galactose, rhamnose, and arabinose with acidified sugars, suchas glucuronic acid and galacturonic acid, to form a random, amorphous heteropoly-mer. Xylose predominates in hardwood hemicelluloses (xylans), while softwoodhemicelluloses (mannans) contain more mannose. The variation in the amount andchemical composition of the hemicellulose in a wood feedstock directly affectsthe in-process properties of its pulp, and all measures of the strength of the fibrenetwork it forms [264].In particular, the hemicellulose component of a fibre system acts to interlinkcellulosic chains with important consequences for the mechanical strength of thenetwork, particularly its tensile strength [284, 285, 290, 309]. Studies have shownthat abundant hydroxyl groups secure these fibre-fibre linkages by means of inter-fibre hydrogen bonds. Thus, a strategy seeking to control a process for end-productmechanical strength must measure the chemical composition of pulp fibres.865.1.2 Pulp Morphology and Its Response to RefiningMorphological factors, such as fibre density, surface area, and the topology of fibre-fibre contact also serve to regulate the effect of a fibre composition profile on thestrength fibre-fibre bonding and the mechanical properties it yields [149, 156]. Therefining or beating of a pulp forces wet fibres between a stationary metal plate, anda moving metal rotor. The mechanical and hydraulic forces associated with refiningalter the physical characteristics of individual fibres and the network they form.Refining removes part of the fibre wall leading to delamination, fibrillationand swelling. This redistributes hemicelluloses from the fibre interior to the sur-face. Fibrillation increases the flexibility of individual strands, and thus makes thenetwork softer. This also increases the exposed surface area available to form hy-drogen bonds, and so adds strength. In addition, interlinking strands incorporatesmore fibre covalent bonds in determining the mechanical strength of the network,as measured by its tensile and bursting strength. However, the process of refiningcan shorten and weaken individual fibres and this reduces the tear strength of asheet.By and large, pulp manufacturers supply raw material as unrefined market pulpto papermakers and other end-product manufacturers, who refine it to fit specificpurposes. The value of a market pulp depends on the physical and mechanicalproperties it can attain in the form of a final product. Target specifications and theirpriority vary with the intended application, but in most cases, buyers seek pulpsthat afford high tensile, burst and tear strength.In order to refer to the physical and mechanical potential of a pulp a marketpulp manufacturer must perform tests on representative portions, refined to specificdegrees in a laboratory pilot plant. Measurements of this kind add to the overallproduction cost of a pulp, and provides information, which might well have beenof use in process optimization, only after production is complete.5.1.3 Prediction of Paper Properties from the Analysis of PulpComposition and MorphologyThe chemical and morphological character of a pulp feedstock varies with the bi-ological characteristics of the raw material, including, for example, the species of87tree, the age and growth history of the stand, and degree of insect infestation, alongwith factors relating to harvest and processing. A paper manufacturer tunes pro-cess parameters to target product specifications and fitness for purpose criteria, butmust contend both with determinative factors and the natural variance of biologicalsystems.Industrial laboratories often use wet chemical analysis and physical tests to de-termine the properties of an in-process pulp and gauge its fitness for a particular enduse [83]. However, laboratory tests require time and the efforts of highly skilledpersonal. The cost and inopportune timescale of such measurements prohibit theiruse for real-time process control.Spectroscopic methods have shown promise as a tool for understanding thestructure of cellulosic pulps. Research has calibrated pulp infrared, near-infraredand Raman spectra to predict many properties normally determined by wet chemi-cal analysis, as well as network physical properties, such as tensile and tear strengths[4, 51, 52, 91, 340, 345, 349].Unlike wet chemical methods, spectroscopic probes can function as online orat-line without the need for sample preparation. However, despite their provenutility for predicting end-product physical and morphological properties in the lab-oratory and the important utility of this information, spectroscopic methods haveyet to show sufficient reliability in the variable environment of an industrial pro-cess line to justify the high technological and dollar cost of entry associated withtheir widespread adoption as a routine tool for process control in the pulp and paperindustry.A significant hurdle exists in the nature of spectroscopic data itself. The amountof information contained in a typical spectrum greatly exceeds the number of cal-ibration samples in a typical data set. This increases the sensitivity of a predictionmodel to noise and reduces the robustness of a calibration owing to the likelihoodof over-fitting.Feature selection methods exist that can remove irrelevant information in aspectrum and reduce the dimensionality of the calibration space. Among manydifferent possible feature-selection strategies, methods based on learning strate-gies, such as a genetic algorithm, offer the best probabilistic, non-local methodsof finding significant features in the space of all possible subsets in a reasonable88amount of time [22]. Here we show that a data-driven genetic algorithm feature se-lection adapted specifically to the problem of optimizing the prediction accuracy ofa spectrochemical classification model can use the variance of the data as a guide todetermine an optimum feature-selection filter for the prediction of paper propertieson the basis of the Raman spectra of cellulosic pulps.Our approach starts with a discrete wavelet transform, which begins to reducethe dimensionality of the spectroscopic information by removing uninformativebaseline variations and high-frequency noise. We then refer to the variance to builda template that guides the process of feature selection in a space of informativewavelets. This filtering tool, which we term a template oriented genetic algorithm,enables the construction of robust calibration models that vastly improve the powerof multivariate analysis to classify vibrational spectra.By isolating robustly correlated wavelet features TOGA also provides a spec-trum of the covariance for chemical interpretation in terms of functional groupfrequencies. We show how the application of TOGA to the Raman spectrum ofunrefined pulps enables the robust prediction of pulp network physical and me-chanical properties for any arbitrarily specified level of pulp refining.5.2 Experimental methods5.2.1 SamplesThis study draws from a supply of 109 northern bleached softwood kraft (NBSK)pulps produced by kraft mills in central interior of British Columbia. These pulpswere manufactured from chip mixtures containing varying amounts of white spruce(Picea glauca), lodgepole pine( Pinus contorta) and sub alpine fir (Abies las lo-carpa). Before refining, these pulps exhibited Canadian Standard Method (CSM)freeness in a range of 640-730 mL as determined by TAPPI standard method T 227om-99.We reserved representative samples of each unrefined pulp from which to fash-ion handsheets for spectroscopic and physical-mechanical characterization. Thepulps were then individually refined over an applied energy interval from 0 to 280kW hours ton−1. Regular periodic sampling and testing determined the points at89which each pulp attained CSM freeness values of 600, 550. 500, 450, and 300 mL.The energies to reach these target values of freeness respectively determinedSRE600, SRE550, SRE500, SRE450 and SRE300 for that pulp, noting that higherrefining energies yield smaller values of freeness. We reserved correspondingpulp samples at each target value of freeness to make papers for the determina-tion of freeness-dependent physical-mechanical parameters for that pulp accordingto TAPPI standards.The present study focuses on six different pulp, fibre and paper sheet proper-ties: tensile strength, burst strength, tear strength, specific refining energy (SRE),wet zero span (WZS) breaking length, and dry zero span (DZS) breaking length.Tensile, burst and tear indices refer to properties of a paper sheet made of refinedor unrefined pulp. We measured these quantities for sheets formed of each pulp ateach refining energy using the following TAPPI standard methods: tensile index (t494 om-01), tear index (T 414 OM-98), and burst index (T403 om-02).SRE denotes to the refining energy required to attain a specified freeness, mea-sured by the volume diverted in a draining stream of 1 L. DZS and WZS provide anindirect gauge of single fibre tensile strength. Wet and dry zero span properties ofunrefined pulp were measured in accordance with TAPPI standard method T 231cm-96. We measured all other properties for sheets prepared from both unrefinedand refined pulps at four different stages of refining.This procedure afforded a large standard matrix of mechanical and physicalproperties of 109 unrefined pulps as well as the same 109 pulps refined to 5 differ-ent values of freeness. The samples examined spectroscopically consisted entirelyof sheets derived from the 109 unrefined pulps.5.2.2 InstrumentationWe have recorded the spectra of handsheets prepared solely from unrefined pulpusing a Raman micro-spectrometer. In this instrument, a 120 µm diameter fibrecouples a 785 nm diode laser onto a dichroic beam splitter, which directs this beamdown the optical axis of an Olympus BX-51 microscope equipped with a ×5 ob-jective with a numerical aperture 0.25. A six-around-one bundle of 100 µm opticalfibres collects backscattered light. An Acton 300 mm monochromator equipped90with a 600 groove mm−1 grating disperses the Raman signal onto a Princeton in-struments PIXIS back-illuminated CCD detector. A LabVIEW user interface col-lects the CCD signal and controls a motorized stage that holds the sample. Thestage moves in a reciprocating x,y lateral motion, phased to sweep the focus of theRaman illumination and collection spot over a 1 cm square of the pulp sheet area.For each sample, we collected the Raman scattered light in two regions ofvibration frequency, from 250 to 1500 cm−1 and from 2800 to 3200 cm−1. AllRaman spectra analyzed in this work combine the signals recorded for 5 acqui-sitions in each of these windows. Each such spectrum used an exposure time of30 seconds. Spectra were saved to files by a LabVIEW data acquisition system,and treated separately using purpose-written MATLAB programs that call uponwidely available discrete wavelet transform (DWT) signal processing and multi-variate analysis packages. We use DWT to filter spectra, removing high-frequencynoise and low-frequency baseline variations. This enables us to select features andbuild multivariate calibration models in a data space of lower dimension.5.2.3 Multivariate AnalysisDiscrete Wavelet TransformDiscrete wavelet transform produces a multi-scale representation of a digitized sig-nal by using a sequence of high- and low-pass filters to sort the signal in terms ofthe frequency with which it varies in the wavelength space of the spectrum. Filter-ing divides this information to resolve the signal into a set of subbands [29- 31].Frequencies of importance, corresponding to peak widths in the wavelength- orλ -space of the original spectrum, appear with large amplitude in the DWT decom-position without loss of λ -space position information. Subsampling removes unim-portant information, including the slowly varying background and high-frequencynoise. Errors of prediction can serve as a guide to choose a wavelet basis andchoose the components that best preserve classification information while remov-ing uncorrelated variance [29].For present purposes, we use a Sym5 wavelet filter. Translating and dilatingthis filter, we form a multiresolution projection of each spectrum, consisting of91subbands with a succession of bandwidths, termed details, together with a finalprojection forming a residual, lowest-frequency subband, called an approximation.We remove components that describe the high frequency noise and low frequencybackground. Typically, we start at the full 1024 point resolution of the collectedspectrum and perform six projections, keeping the four detail components fromlevel two to six. We then build partial least squares (PLS) calibration models fortarget properties in the reduced space of the retained wavelet coefficients, and findthe root-mean-square error of prediction (RMSEP) for that property. We selectscales at which to cut the high and low frequencies that minimize the RMSEP.Template Oriented Genetic AlgorithmWe employ a novel step of spectral feature selection to further focus and economizeour multivariate calibration models, applying a genetic algorithm (GA) to establishthe most informative feature subspace. To overcome the well-known disadvantagesof conventional GA, the present method, termed, Template Oriented Genetic Al-gorithm (TOGA) uses a data driven approach [35]. This means that we determinethe significance of each feature first and then decide the optimal combination offeatures. The factor that controls the flow of information is the data itself not apredetermined logic of the program.The core algorithm of TOGA proceeds as follows [35]:1. In a data space defined by the retained DWT coefficients, we calculate theimportance of spectral predictors with respect to a target property accordingto the equationSi =mean(βi)std(βi)(5.1)where mean(βi) and std(βi) represent the mean and standard deviation of thePLS regression coefficients of predictor i, (βi) in 300 different PLS modelsfor the covariance of the Raman spectra in a wavelet basis. We select the 200predictors with largest importance to define the initial predictor space for thetarget property.2. Selecting from this predictor space, we define a population of 50 chromo-somes, each composed of 10 features, i, selected randomly from among the92200 most important predictors, using a probability weighting determined byproabilityi =Si∑ni=1 Si(5.2)3. We use each chromosome to define a unique, feature-selected data matrixrepresenting the reduced Raman spectra of all 109 unrefined pulp samples.For each such data matrix, we randomly select 75 percent of the spectra tobuild a PLS calibration model, and the remaining 25 percent to validate it.Defining the associated RMSEP of validation as an ensemble-error-criterionfitness function for each chromosome, we determine its suitability accordingto the equation:RMSEP =(∑mj=1((yˆ j)− y j)2m)1/2(5.3)where yˆ j and y j refer to the predicted and measured property values of sam-ple j respectively, and m is the number of samples.4. On the basis of this criterion, we select typically 10 of the most fit chromo-somes to pass unmodified into the next generation. Then, applying ensemble-error weighting to the remaining chromosomes, we form 20 fit pairs and ap-ply random single-point crossover to generate two children each. Drawingnew genes from the predictor space, we mutate five of these at random.Using the new chromosomes as templates, we build a set of 50 new PLSmodels, which generates a new set of fitness functions. We repeat this cy-cle through 10 generations, and save the single top chromosome, definedglobally by the smallest ensemble error.5. After 10 iterations of this whole process, or after an inner loop of 10 gener-ations produces no more than a negligible reduction in prediction variance,we examine the genes that appear in each final top chromosome. Choosing10 genes on the basis of appearance frequency, we construct a final chromo-some representing the set of wavelet features selected by TOGA.93We believe that TOGA serves well to addresses weaknesses in previous methodsfor converting the spectra into a lower number of features. For example, conven-tional feature-selection methods do not exploit the functional character of spectra(as established e.g. by wavelet transform) and thus generally represent selectedfeatures by defining a discontinuous set of spectral positions. TOGA, on the otherhand, projects each spectrum on a sub-band template that points to chemical infor-mation in hidden band shape variations.5.3 ResultsThis section surveys the physical and mechanical standardization properties of thefull set of unrefined and refined pulp sheets. It then presents the unprocessed andDWT-filtered Raman spectra obtained exclusively from measurements on the 109unrefined pulp sheets. We follow this by an analysis that explores the degree towhich the spectra of unrefined pulps predict the physical and mechanical proper-ties of sheets formed using pulps subjected to varying degrees of refinement. Toclassify the effect of refining on the physical properties of interest, we employTOGA to isolate the best predictors.5.3.1 Physical and Mechanical Properties of Unrefined and RefinedPulp SheetsIn standard tests, the handsheets investigated in this study yield distributions oftensile and tear indices that depend on the freeness of the primary or refined pulp,as displayed in Figure 5.1. For example, tensile strength increases with refining(decreasing freeness). Pulps with a freeness of 600 mL yield tensile indices thatdistribute about a mean of 6 km with a standard deviation of 1 km, while morerefined pulps, with a freeness of 300 mL exhibit a mean tensile index of 10.5 kmwith a slightly larger standard deviation.Unrefined pulps, however, exhibit the highest tear index. This quantity de-creases as freeness drops for increasingly refined pulps. Thus, for pulps with afreeness of 300 mL, we find a tear index distributed about a mean of 12 mN m2g−1 with a standard deviation of about 2 mN m2 g−1. Less refined pulps with afreeness of 600 exhibit a tear index mean of 20 ± 3 mN m2 g−1.942 4 6 8 10 120510152025303540Tensile strength (km)Number of samples500450300550600UnrefinedStudent Version of MATLAB5 10 15 20 25 30 350510152025303540Tear (mN m2 g−1)Number of samples Unrefined500550600450300Student Version of MATLABFigure 5.1: Histogram showing tensile and tear properties of six different of unrefined and refined pulpsheets, freeness= 600, 550, 500, 450, 300. a) tensile, b) tear.Figure 5.2 cross correlates the tensile strength of pulps at different degrees ofrefinement, as reflected by freeness. The diagonal entries reflect the distributionsof tensile strength within a freeness class, corresponding to the histograms super-imposed in Figure 5.1. Notice that the tensile strengths of pulps with intermediatefreeness correlate strongly. By comparison, however, the tensile strength of an un-refined pulp predicts poorly for the tensile strength after refining. Similarly thetensile strength of a pulp refined to a freeness of 300 mL changes to an extent notwell predicted by the tensile strength of that pulp at intermediate refining.5.3.2 Raman Spectra of Unrefined Pulp SheetsFigure 5.3 shows the raw and DWT Raman spectra for our set of 109 unrefined kraftpulp samples. The raw Raman spectra exhibit weak signals for some samples witha persistent background owing to fluorescence. Note the similar band patterns, butsome variations in relative intensities. After discreet wavelet transform (DWT) pre-processing, the filtered spectra present a flat background with little high-frequencynoise.The region below 1500 cm−1 corresponds to motions of the C-C and C-Oframework, while the broad intense peak in the interval from 2800 to 3000 cm−1represents to O-H stretching. We can associate regions labeled F and G in Figure5.3 with modes that involve considerable coupling of methyl bending, methylene95Figure 5.2: The correlation between tensile strength of six different of unrefined and refined pulp sheets,freeness= 600, 550, 500, 450, 300.rocking and wagging and C-OH in-plane bending.The peaks in 1090, 1120, and 1150 cm−1 (region E) represent modes involvinganti-symmetric C-O stretch and ring vibrational modes. We can assign the peakat 897 cm−1 (region C) to HCC and HCO bending localized at C(6) position inglucose. Region A, below 600 cm−1 relates predominantly to skeletal, CCC, OCO,COC, and OCC bending vibrations. The particular feature at 477 cm−1 (region B)reflects the amount of hemicellulose in the pulp.5.3.3 TOGA-Selected Sub-Spectra Associated with ParticularPhysical Properties of Unrefined Pulp SheetsFigure 5.4 shows spectra constructed from a TOGA feature selection targeted to aselected set of physical and mechanical properties of pulp sheets.Compare the TOGA reconstructed spectra related to the tensile and burst strength96250 550290027001750F GEDCABHRaman shift (cm-1)850 1150 1450(a)250 550290027001750F GEDCABHRaman shift (cm-1)850 1150 1450(b)Figure 5.3: Pulp Raman spectra before/after preprocessing by discrete wavelet transform with selectedwavelet range of (2-6) a) Raw Raman spectra, b)DWT Raman spectra.of unrefined samples plotted in Figure 5.4(a) and (c). Note that TOGA isolates sim-ilar spectral regions for both properties; however, TOGA selects features in eachwavenumber interval that have different intensities. These TOGA reconstructedspectra indicate that features in regions B and C have greater importance in multi-variate models for tensile strength, while features in region F play a more important97(a) (b)#(c)   3BNBO4IJGU	DN (d)(e) (f)Figure 5.4: Pulp Raman spectra after discrete wavelet transform and TOGA feature selection guided tooptimize the following properties of unrefined pulp a) tensile strength, b) SRE600, c) burst strength, d)dry zero span, e)tear strength, f) wet zero span.role in classifying for burst. Note that region B appears with varying intensity inthe TOGA spectra coding for tensile, burst, and tear.The TOGA spectra displayed in Fig. 5.4(e), selected for the prediction of tearstrength, give greatest weight to region H. Raman features in region A, E, and Falso appear to be important for the classification of tear strength, while structure inother regions of the spectrum seem unrelated to this property.The TOGA spectrum optimized to predict SRE600 shows a simple pattern with98strong features at 378 (A) and 1095 (E) cm−1, and weaker structure in region G.Figure 5.4, frames (d) and (f) show TOGA reconstructed spectra for dry-zero-span (DZS) and wet-zero-span (WZS) breaking strengths. Here we see that featuresat 379 cm−1 (A) and 1095 cm−1 (E) dominate in modeling zero span strength as aclassification property in both cases. TOGA marks region B as an important featurefor WZS and not for DZS.5.3.4 TOGA-Selected Sub-Spectra Associated with ParticularPhysical Properties of Refined Pulp SheetsWe examine effect of refining by classifying the spectra of unrefined pulps for theproperties they exhibit at lowered freeness. Figure 5.5 show how TOGA-selectedfeatures change depending upon the targeted freeness.(a)     Raman Shift (cm-1)250         850        1450       2050        2200       2800(b)     Raman Shift (cm-1)250         850        1450       2050        2200       2800(c)     Raman Shift (cm-1)250         850        1450       2050        2200       2800Figure 5.5: Raman spectra of unrefined pulps after wavelet transform and TOGA preprocessing to opti-mize the prediction of burst strength, tensile strength, tear strength and SRE600 at freeness of (a) 600ml, (b) 450 ml and (c) 300 mlJust as with Raman spectra, feature-selected on the basis of their unrefinedproperties, region H appears prominently as a marker for tear strength comparedto other properties at a freeness of 600. TOGA also singles out regions A and Gor F as a critical features for all properties at this degree of refining. We note that99TOGA points to region B as an essential feature for tensile, burst, and tear.This pattern changes dramatically for freeness 450 in a way that we see re-peated in spectra for all properties. TOGA chooses the wavelet that forms thefeatures in region E and targets wavelet position on the shoulders of other fea-tures. For tear, we see that the importance of region H decreases significantly withdecreasing freeness.For refining energies that produce a freeness of 300 ml, a TOGA model target-ing tensile and burst calls for a decrease in the weight assigned to region E, with are-emergent significance in regions A, B and F, together with an outsized emphasison structure in region H, all regulated by fine tunning the selected position of thewavelet. TOGA reconstructed spectra for tear at a freeness of 300 ml resemblesthat targeted to freeness 450 emphasizing regions E, C, and A. Reconstructed spec-tra for SRE display a simple pattern at all refining energies, dominated by waveletsin region E.5.4 DiscussionThis work explores the extent to which TOGA feature selection, targeted to theprediction of pulp sheet properties at specific degrees of refining, yields calibrationmodels that can predict a pulp sheet property at any refining energy from the Ramanspectrum of the pulp before refining. To understand the basis on which one mightmeet such a classification objective, we begin with an examination of the propertiesof pulp fibres and their network effect on physical properties of raw pulp.We then discuss how refining changes the physical chemical nature of pulpfibres, and the way these changes alter the fibre network to affect its physical andmechanical properties. Much of this discussion will focus on fibre-fibre hydrogenbonds and how increases in these interactions predict increases in the tensile andburst strength of a pulp sheet, while anti-correlating with tear strength.5.4.1 Pulp Properties as Affected by Fibre Structure and Fibre-FibreInteractionsThe physical and mechanical qualities of a paper sheet depend both on the natureof individual fibres, and characteristics of the fibre network determined by fibre-100fibre bonding interactions. Refining affects these molecular properties in two ways.Mechanical abrasion breaks fibres and disrupts natural structure to yield fibrillatedfibres of shorter length. Refining also acts on fibres to affect the crystallinity of thecellulose they contain [101].Tensile and burst strength depend on the network strength of fibre-fibre interac-tions. Refining increases the strength of the network by producing shorter, openedfibres, which exposes a larger surface area, increasing the number of hydrogenbonds. At first, these interactions increase cellulose crystallinity and strengthenthe network [53]. On the other hand, tear strength and SRE depend for the mostpart on the length and strength of individual fibres. Because refining shortens fi-bres and disrupts fibre structure, these qualities decreases with increasing refiningenergy.Figure 5.1 summarizes trends in the distribution of physical properties of sample-set pulp sheets with different refining energies by plotting the distributions ofmeasured properties as histograms for pulps refined to various levels of freeness.Consider the tensile histograms. For intermediate refining energy, they exhibit aslightly right-skewed distributions. High refining energy, which results in a free-ness of 300 mL gives rise to a plateaued distribution, which looks like that of theunrefined pulp.Changing the distribution pattern from plateaued to right-skewed and back toplateaued reflects two distinct changes that occur in the structure of the pulp as therefining level increases from intermediate to high. We examine Raman spectra foran evidence of these chemical and morphological changes in fibres and fibre-fibreinteractions at these different levels of refining.Figure 5.2 shows the correlation between values of tensile strength measuredfor given pulps at levels of refining characterized by freeness values from 600 to300 mL. Here we see that the tensile strength of the unrefined pulp is a poor pre-dictor of tensile strength at any level of refining. Similarly, the tensile strength of ahighly refined pulp (300 mL) predicts poorly for the tensile strength at lower levelsof refining. But, values of tensile strength correlate strongly between pulps refinedto intermediate freeness in the range of 600 and 450 mL.This suggests again that fibre and fibre-network variations at the molecularlevel produced by refining abruptly change in nature as refining decreases the pulp101freeness from 450 to 300 mL. In the following section, we will consider how re-fining affects the properties of individual fibres and their interactions in a network,as manifested by changes in the vibrational spectrum. We then look at the trendsobserved in vibrational structure revealed by TOGA feature selection, chemomet-rically targeted to the prediction of pulp properties at various levels of refining.5.4.2 Vibrational Manifestations of Pulp Fibre and NetworkMolecular PropertiesBand positions and intensities of certain vibrational features in the Raman spectrumof a pulp vary to reflect the degree of hydrogen bonding and cellulose crystallinity.As indicated in Figure 5.3, region H (2800-3000 cm−1) reflects hydrogen bondswhich involving principally 6-3´ OH...O interfibre bonding and intrafibre bondingfor 2-6 OH...O and 3-5 OH...O [224].Intra-fibre hydrogen bonds determine fibre extensibility, which affects fibrelength and tertiary structure. Interfibre hydrogen bonds join fibres together. Thenumber of the intermolecular hydrogen bonds thus determine the strength of the fi-bre network. The intensity and the shape of features in region H reflect the numberand the strength of hydrogen bonds that define cellulose crystallinity, determinethe flexibility of fibres, and regulate the fibre surface area in molecular contact.The spectroscopic region below the 1500 cm−1 represents vibrational modesof the cellulose backbone, which respond with particular sensitivity to the degreeof cellulose crystallinity. For instance, features in region G (1474 cm−1) shift tolower frequency and broaden with a reduction in cellulose crystallinity, moving to1462 cm−1 for fully amorphous cellulose [266]. Region E displays ring vibrationalmodes involving skeletal stretching. The spectrum in this region varies with theorientation of glucose linkage in the cellulose chain. Previous study has shownthat tensile deformation shifts the 1095 cm−1 band in region E to lower energy,while pulp fibre strain affects the intensity ratio of bands at 1120 and 1090 cm−1[340]. Conventional assignments relate the maxima at 379 (A) and 1090 (E) cm−1to the cellulose crystallinity [266, 340], while the intensity and the position ofpeaks located at region C reflect the degree of cellulosic disorder and the size ofcellulose crystallites [266, 340].The strength of a fibre network depends on the strength of individual fibres102as well as the strength and distribution of bonds between fibres. The fraction ofcrystalline cellulose, the size and orientation of crystallites, and the geometry ofthe crystal lattice all have an effect on the physical and mechanical properties of apulp fibre network [344]. Raman spectroscopic information highlighted by TOGAin regions C, E and G bear on these factors regulated by crystallinity.To explore how factors related to fibre structure and hydrogen bonding influ-ence the properties of unrefined pulp, we provide the following analysis of Ramanstructure isolated by TOGA feature selection geared to predict various differentphysical and mechanical properties of pulp networks.5.4.3 TOGA-Selected Sub-Spectra Provide Evidence forCompositional Factor Underlying Pulp Physical Properties ofUnrefined Pulp SheetComparing the TOGA reconstructed spectra from the analysis of unrefined pulp(5.4), we see that TOGA selects simple patterns of features for parameters thatdepend on the properties of individual fibre such as DZS, WZS, and SRE600. Ityields a more complex pattern containing more features for properties dependingon intra-fibre bonding. Studying TOGA reconstructed spectra for the unrefinedpulps suggests that features in regions A and E classify most directly for zero-span and SRE600. Conventional measures the ratio of these two peaks reflects thecrystallinity of cellulose, we can therefore conclude that fibre crystallinity correlatewith WZS, DZS, and SRE600.Similarly, we can recognize that the hemicellulose (region B) appears signif-icantly in the TOGA model for every parameter that depends on fibre networkstrength, such as tensile, tear, and burst. This region is absent for WZS, which de-pends only on the property of single fibres. Although SRE600 depends for the mostpart on the characteristics of individual fibres, the hemicellulose region (region B)appears in models for this property. The following sections detail the essentialRaman features associated with each property of interest.Strength properties: Tensile and BurstTOGA suggests that both the structure of the cellulose (region C and F) and a facil-ity for fibre-fibre attractions (region H) affect tensile and burst strengths. However,103regions that correspond to crystallinity of the cellulose appear with higher inten-sity (TOGA centers selected wavelet on key feature). TOGA chooses region B,assigned to hemicellulose as a particularly informative feature for these two prop-erties. We find this association with hemicellulose content strongest for tensile.Tensile strength directly relates to the hemicellulose features because the presenceof this substance increases the bonding interactions between fibres [139].TearTOGA indicates that tear strength of unrefined pulp depends on both on fibre mor-phology and the strength of interfibre bonding. Region H signifying the degreecorrelated most significantly hydrogen bonding with this property. The extensibil-ity of hydrogen bonding depends on the fibre length.Specific Refining Energy 600TOGA selects regions A, E, and G indicating the importance of the cellulose struc-ture for SRE600. Fibres with higher cellulose crystallinity, signified by the am-plitude of features A and E, require more energy to break structural units. TOGAselects region H, correlating to hydrogen bonding, and region B reflecting hemi-cellulose. These attributes affect a pulp’s ability to hold water and consequentlymake the process of draining slower.Strength Properties of Single Fibre: Wet and Dry Zero SpanSpectra reconstructed by TOGA for DZS and WZS present dominant peaks in re-gions E and A related to the crystallinity and structure of the fibres. TOGA recon-structed spectra show small differences in selecting hemicellulose signal (regionB) as important feature for dry zero span. Selecting region B for DZS by TOGArepresents the effect on interfibre bonding and bridging of hemicellulose on dryzero span strength. Researches concluded that dry-zero span strength changes withinterfibre bonding [36], while wet zero span effectively removes all bonds. Hemi-cellulose and fines also can bridge weak section of fibres for dry sheets.1045.4.4 TOGA-Selected Sub-Spectra Provide Evidence forCompositional Factors Underlying Pulp Physical Properties ofRefined Pulp SheetRefining makes fibres softer, shorter, more flexible, and gives them higher surfacearea. With higher beating, fibres become more flexible, and the interaction be-tween fibres increases. Results suggest that during the refining process, the func-tional group characteristics of the pulp remains unchanged, while the number andthe strength of hydrogen bonds increases [101]. The crystallinity of the fibre isalso affected by beating [137].We examine features selected by TOGA geared to aprediction of properties of the refining to understand the spectroscopy factors thatpredict for the magnitude of these changes on the several properties of pulpsheetsusing Raman spectra.Freeness 600Figure 5.5 (a) shows TOGA reconstructed spectra targeting burst, tensile, tear, andSRE600 for a hypothetical freeness of 600 ml. TOGA shows the similar complexpattern of features when predicting for the unrefined pulpsheet and one with a free-ness of 600 ml. The similarity in the TOGA patterns for unrefined and refinedpulp with freeness 600 can be related to the range of freeness of the unrefined pulpwhich is in the range of 638-704 with the average of 680. We don’t expect to seethe dramatic change in the characteristics of pulp fibres and paper network by de-creasing freeness from 680 to 600. The TOGA spectra targeting these parametersindicates that the hydrogen bonding (region H) seems to respond most as an indi-cator of the effect of refining on tear, while cellulose crystallinity (region G and F)seems to code most for tensile and burst. The ratio of intensities E/A points with ahigh degree prediction value to the effect of refining on SRE600. TOGA choosesregions A, E, and G when predicting SRE600. All these features signal the struc-ture and the crystallinity of the cellulose fibre. TOGA reconstructed spectra forother properties combine features related to both fibre structure and the networkof fibres. TOGA selects regions referring to hydrogen bonding and hemicellulosefor all the properties of lightly refined pulp, reflecting the influence of fibre-fibrebonding on all of these quantities.105Freeness 450The pattern of TOGA reconstructed spectra configured to predict for a freeness of450 shows interesting changes. Here, in every case TOGA emphasizes region E,which represents the crystallinity of fibre. Previous work suggests that the processof beating decreases cellulose crystallinity in the beginning, but increases it withan increase in beating time [137]. TOGA underlines this result by singling outcombination of features attributed to cellulose crystallinity. TOGA also selects lowintensity peaks associated with hydrogen bonding for all properties except SRE-450. This omission suggests that single fibre distinguishes SRE450, while bothsingle fibre and network properties operate to determine tensile, burst, and tear ata refining energy that gives a 450 ml freeness. TOGA suggests that hemicellulosefeatures play an essential role in predicting tensile and burst of refined pulpsheetswith a freeness of 450 ml pointing to the important influence of fibre-fibre bondingon these properties.Freeness 300TOGA wavelet patterns suggest that the physical characteristics of single fibresmust affect on the mechanical properties of unbeaten pulp or pulp with low beatingenergy. Single fibre effects diminish for pulp beaten to further freeness of 300 ml.Here, for tensile and burst fibre interactions became the dominant factor. Unrefinedpulp contains long cells with thick walls which supports limited low fibre- fibrebonding. Refining collapses fibres and make them more flexible. The presence ofthe shorter fibres and fines increases the interaction between fibres. Moreover, theflexibility of fibres enables them to form configuration of increased contact area.This larger bonding surface supports a greater number of hydrogen bonds. TOGAconfirms this idea by pointing to the hydrogen bond region as a significant predictorof tear and burst for highly beaten pulps with freeness of 300.In contrast, for tear TOGA reconstructed spectra at varying freeness implythat the hydrogen- bond region becomes less important. As we mentioned above,greater fibre- fibre bonding improves all physical properties but tear strength. Thetear index so determined by the characteristics of single fibres balanced with thenetwork strength between fibres. For unrefined pulps and pulps treated with low106refining energies, energy applied out of the fibre plane pulls the pulpsheet networkapart. TOGA recognize this in emphasizing the hydrogen bonding region H inmodels for all strength attributes. The significance of hydrogen bonding for ten-sile and burst strength at higher refining energy increases with the formation ofshorter more flexible fibres and fines, as evident in figure 5.4-c. Hydrogen bondingbecomes less predictive for tear strength with increased refining. The presence ofhemicellulose adds to the hydrogen bonding capacity of a pulp, and TOGA iden-tifies similar predictors of fibre network strength in region B of the pulp Ramanspectrum.On the other hand, tensile and burst strength depend on fibre strength (crys-tallinity), in combination with fibre-fibre hydrogen bond strength along with therelative contact area. Refining produces shorter fibres and produces fibre frag-ments known as fines. These fibre fragments fill the space between binding longfibres. This increases the number of hydrogen bonds and fibre-fibre area in contact.TOGA reconstructed spectra geared to predict tensile and burst emphasize the hy-drogen bonding region. The general linking predicted increases these parametersto the growing number of hydrogen bonds with increased refining.5.4.5 Multivariate Classification of Pulp Raman Spectra forEnd-Product Properties Using Template Oriented GeneticAlgorithmWe have investigated the effectiveness of Raman spectroscopy as a means of clas-sifying a pulp to a degree that accurately predicts the properties of the productsheet forms in advance of its manufacture. To begin, we take the Raman spectra ofunrefined pulp sheets. We apply TOGA to economize the Raman spectral represen-tation. We evaluate the effectiveness of spectroscopic measurement combined withchemometric analysis by constructing prediction models and testing prediction ac-curacy using the method of partial least squares (PLS) reserving with independentpretreated data subsets as validation standards.Figure 5.6 and 5.7 illustrate PLS models for 4 properties before and after apply-ing TOGA. We have made PLS prediction model by calibration the Raman spectraof unrefined pulps against properties of unrefined and refined pulp with five differ-ent freeness (600, 550, 500, 450, 300). PLS models for tensile and burst shown in1070 2 4 6 8 10 12024681012Measured tensile strength (km)Predicted tensile strength (km)Student Version of MATLABa)Tensile0 2 4 6 8 10 12024681012Measured tensile strength (km)Predicted tensile strength (km)Student Version of MATLABb)Tensile−TOGA0 1 2 3 4 5 6 7 8 9 10012345678910Experimenta burst (kPa m2 g−1)Predicted burst (kPa m2  g−1 )Student Version of MATLABc)Burst0 1 2 3 4 5 6 7 8 9 10012345678910Experimental burst (kPa m2 g−1)Predicted burst (kPa m2  g−1 )Student Version of MATLABd)Burst−TOGAFigure 5.6: PLS models for untreated and treated Raman spectra by TOGA with latent variable equals to7 a)Tensile (untreated Raman) b) Tensile (treated Raman) c) Burst (untreated Raman) d) Burst (treatedRaman).Figure 5.6 show that TOGA improves the prediction models based wholly on spec-tra of unrefined pulp for the application of these properties at any level of refining.TOGA narrows the scatter for both the training and validation data sets.We should expect to achieve better a PLS model for unrefined properties be-cause we used unrefined pulp sheet; however, Figure 5.6 shows the similar or evenbetter correlation for higher refined properties. One reason for this observation canbe relied on the weakness of Raman in detecting effective factors; which becomeless important for properties of refined sample. One example of these factors canbe fibre length, width, and cell wall thickness. For properties of unrefined sample,1080 5 10 15 20 25 30051015202530Experimental tear (mN m2 g−1)Predicted tear (mN m2 g−1 )Student Version of MATLABa)Tear0 5 10 15 20 25 30051015202530Experimental tear (mN m2 g−1)Predicted tear (mN m2 g−1 )Student Version of MATLABb)Tear−TOGA0 50 100 150 200 250 300 350050100150200250300350Experimental specific refining energy (KWhr t−1) Predicted specific refining energyStudent Version of MATLABc)SRE0 50 100 150 200 250 300 350050100150200250300350Experimental specific refining energy (KWhr t−1)Predicted specific refining energy (KWhr t−1 )Student Version of MATLABd)SRE−TOGAFigure 5.7: PLS models for untreated and treated Raman spectra by TOGA with latent variable equals to 7a)Tear (untreated Raman) b) Tear (treated Raman) c) SRE (untreated Raman) d) SRE (treated Raman).single fibre properties is the dominant factor in predicting them, as we discussedabove. Since Raman has limitation in detecting this property, PLS model usingRaman fails to predict unrefined properties precisely. On the other hand, by in-creasing beating the effect of fibre length and width become less important whileinterfibre bonding shows the most essential impact on refined properties. Ramanspectroscopy is able to detect hydrogen bond, which is the dominant interfibre bondin paper samples. So, we achieve more reliable PLS model using Raman spectrafor refined properties.Figure 5.7 shows that PLS models that predict tear and SRE for untreated Ra-man spectra and treated Raman spectra by TOGA. PLS models for tear and SRE109follow much the same pattern as models for tensile and burst. TOGA improves theregression coefficient and correlation for both characteristics. However, contrast-ing with PLS models for tensile and burst, here we see poorer prediction occurs forthe higher tear indices and SRE observed in unrefined pulps. We notice that TOGAsubstantially improves the quality of prediction models of these attributes for pulpsafter refining.We report error of prediction for all properties for untreated and treated Ra-man spectra in table 5.1. We observe that untreated Raman spectra yield highrelative errors of prediction for almost all the physical properties. Weak Ramansignal, overwhelming-fluorescence background, matrix interference, and the overdetermination associated with a high numbers of features all contribute to this lim-itation of untreated Raman spectra. We note significant improvements in relativeprediction error after selecting important Raman features by TOGA.This confirmsthat TOGA serves well to filter out uncorrelated features without removing impor-tant information. This improvement is more noticeable for parameters in unrefinedsamples. Table 5.1 also shows that Raman predicts all properties except tear withhigher beating energy and lower freeness more accurately.Refining collapse fibres, after refining they becomes more flexible and theirbonding surface area is increased. By increasing the refining energy and havingsmaller fibres, fibres interact more, detecting hydrogen bond interactions thus be-comes more important, improving the discerning power of Raman analysis. Thisobservation implies that Raman spectroscopy better predicts the physical proper-ties of samples with shorter fibres because it effectively detects the potential forhydrogen bonding interactions in the sample.5.5 ConclusionRaman spectroscopy has particular strengths suited to pulp and paper analysis. Itextracts chemically specific signatures from aqueous samples with little or no sam-ple preparation. Here we present new computational routines that enable multivari-ate analysis to overcome potential limitations of sample fluorescence and substan-tial uncorrelated variance. TOGA classifies unrefined pulp spectra to a degree thataccurately predicts the attributes of finished paper. TOGA provides information110for any subsequent degree of pulp refinement that relates morphological detailsof pulp structure and bonding to chemical and physical properties of finished pa-pers. Technology emerging from this research will ultimately point the way to anon-line process control sensor that can measure production pulp stream chemicalcomposition and gauge its conformance with end-use paper sheet specifications.111Table 5.1: Prediction error for all physico-mechanical properties of pulp samples using treated and un-treated Raman spectra.Parameters Range Mean Raman-DWT-PLS Raman -DWT-TOGA- PLSTensile (initial) 2.50 - 4.87 3.72 0.61 0.39Tensile 600 4.30 - 7.67 6.17 0.68 0.49Tensile 500 6.50-9.63 8.24 0.73 0.57Tensile 450 7.90-10.63 9.34 0.81 0.61Tensile 300 9.20-11.56 10.32 0.62 0.52Burst (initial) 1.36-3.37 2.25 0.40 0.24Burst 600 3.2-5.95 4.38 0.41 0.31Burst 500 4.90-7.63 6.21 0.62 0.41Burst 450 5.76-8.37 7.17 0.75 0.52Burst 300 6.55-9.18 8.01 0.69 0.49Dry zero span 14.42-16.78 15.51 0.97 0.76Wet zero span 12.40-15.82 14.42 0.80 0.64Tear (initial) 19.86-30.34 25.83 2.81 1.92Tear 600 17.34-24.29 20.24 2.15 1.37Tear 500 12.15-20.13 15.53 2.09 1.41Tear 450 9.78-18.37 13.25 1.56 1.09Tear 300 8.25-15.22 11.44 1.61 1.12SRE600 23.91-82.73 52.19 8.40 5.73SRE 500 60.81-162.50 113.77 15.09 11.66SRE 450 92.35-217.97 160.16 26.95 22.35SRE 300 152.90-309 239.85 31.91 27.15112Chapter 6Combination of MultipleSpectroscopy Techniques UsingData Fusion for EnhancedPrediction Modeling ofPhysical-Mechanical Propertiesof PaperWe demonstrate that fusing the Raman, NIR, and ATR-FTIR spectra of unrefinedpulps takes advantages of the synergistic effect of the information acquired bythe three vibrational spectroscopic techniques to accurately predict the mechani-cal properties of the corresponding final paper product. We investigate strengthsand weaknesses of each technique when applied individually to predict variouspulp properties. We compare low-, mid-, and high- level data fusion for improv-ing the prediction of paper properties of interest. For mid-level data fusion, wemerge selected variables of Raman and NIR using a powerful chemometric featureselection technique termed Template genetic Oriented Genetic Algorithm (TOGA)for Raman and a well-known feature reduction method referred to as Orthogonal113Signal Correction (OSC) for NIR. For high-level data fusion, we combine the pre-diction results obtained individually for each spectroscopic measurements. Theresults show generally that data fusion improves the efficiency of prediction mod-els in predicting physical and mechanical paper properties, compared to modelsbuilt using only individual techniques.6.1 IntroductionThe spectroscopic analysis of pulp and paper has garnered significant interest re-cently as a robust, rapid, and cost-effective alternative to the traditional wet-chemicalmethods traditionally employed by industry. Spectroscopic techniques show inher-ent sensitivity to many properties of interest, but do not yield information with thesame clarity as wet-chemical methods and thus must be paired with advanced datatreatment techniques in order to yield useful results.Three spectroscopic techniques have shown particular promise: near-infraredspectroscopy (NIR), attenuated total reflection fourier-transform infrared spectroscopy(ATR-FTIR), and Raman spectroscopy. Besides being rapid and cost-effective,these spectroscopic techniques also present the distinct advantage of being en-tirely non-destructive, unlike wet chemical processes, and can be adapted for on-line process monitoring applications. All have been investigated to some extent[44, 63, 217, 241, 251, 315, 320, 349]. Each presents strengths and weaknesses inpredicting pulp and paper properties, and these must be considered when choos-ing a technique best suited for a particular application or material property to beinvestigated. The ability to determine the cellulose crystallinity of the pulp is ofparticular interest, since the crystallinity plays a role in predicting bulk pulp prop-erties.NIR spectroscopy has been used in a variety of contexts within the pulp andpaper industry. Antti et al. combined NIR with partial least-squares (PLS) mod-eling to identify tree species contributing to a pulp mixture, and predict a varietyof pulp properties with a high degree of accuracy [112]. Hauksson et al. pub-lished a similar study in 2001 in which they used nearly identical techniques to114accurately predict various wood meal properties [120]. Several more studies havebeen published studying the Eucalyptus genus, in which the authors were ableto accurately predict pulp yield from wood, as well as various pulp properties[74, 89, 211, 241, 242, 251, 252, 269, 270]. Cooper et al. published a study in2011 that reviewed the applications of NIR spectroscopy to the characterization ofwood and pulp products in some detail. The original research they reported fo-cused on the moisture content of wood samples, as well as species determination.[65] NIR spectroscopy has been demonstrated to be sensitive to many of the com-ponents of wood and pulp such as lignins, cellulose, and hemicellulose, as wellas the interactions between them [120, 178, 349]. When combined with process-ing techniques such as PLS, NIR spectroscopy is generally a very robust methodfor constructing accurate prediction models for wood and pulp properties. Nelsonand O’Connor introduced the measure Total Crystallinity Index (TCI), which is theratio of near-infrared absorption bands at 1372 cm−1 and 2900 cm−1. They alsoestablished an empirical “crystallinity index” as the ratio of near-infrared absorp-tion bands at 1429cm−1 and 893 cm−1, implying that NIR spectroscopy offers agood monitor of pulp crystallinity [217].Other spectroscopic techniques, such as ATR-FTIR and Raman spectroscopies,may be more sensitive to certain aspects of wood and pulp products than NIR spec-troscopy, and thus may be better suited for specialized applications. ATR-FTIRspectroscopy is a surface-probing technique, and thus is more sensitive to surface-related properties [349]. On the other hand, Raman spectroscopy is almost com-pletely insensitive to water, thus avoiding some of the overwhelming water signalsobserved in infrared spectra.ATR-FTIR spectroscopy has been shown to be a useful technique for predictingpulp properties, although is not as prevalent in the literature as NIR spectroscopy.Numerous studies have been published detailing its use along with PLS and otherchemometric techniques. Many have used ATR-FTIR spectroscopy to probe phys-ical properties or changes such as surface degradation [138, 184, 206], surfacecoatings [90, 205, 207], thermal treatment [130, 343], and species differentiation[130, 360]. Our own unpublished work indicates that ATR-FTIR spectroscopy115is much more sensitive to surface-dependent properties such as burst index, lightscattering, and SRE600, but is less sensitive for bulk properties such as wet zerospan and freeness. ATR-FTIR spectroscopy can detect changes in the crystallinityof cellulose, which are seen by movement of cellulose peaks that occur at 1430,1163, and 897 cm−1 [44, 63, 217, 218, 232], as well as the peak at 3330 cm−1. Thelatter is assigned to the O – H stretching band representing the specific O(3)H –O(5) intramolecular hydrogen-bond. It has been shown that this peak shifts to3420 cm−1 for amorphous cellulose [162].Raman spectroscopy, as applied to characterizing wood and pulp samples, isnot as widely researched as NIR and FTIR. Although Raman spectroscopy avoidsinterference due to the presence of water in samples, it is far more sensitive tofluorescence interference. Although Wiley and Atalla published a study detailingband assignments of the Raman spectra of cellulose [340]. However, pulp andwood systems are much more complex and often require more advanced chemo-metric processing techniques [91, 265, 321, 349]. Despite these complications,Raman spectroscopy has been used to measure the lignin content of pulp samples[138, 302, 326, 349], as well as cellulose crystallinity [265, 266, 320, 321]. Thefeature at 1474 cm−1 in the Raman spectrum of cellulose shifts to a lower fre-quency and broadens with a reduction in cellulose crystallinity, appearing at 1462cm−1 for amorphous cellulose. The peak at 897 cm−1, which is assigned to HCCand HCO bending localized at the C(6) position in glucose, changes both in fre-quency and intensity to reflect the degree of disorder in cellulose and the size ofthe regions of crystalline cellulose.Each of these spectroscopic techniques has its own advantages and disadvan-tages for predicting a variety of pulp properties. It stands to reason that we canimprove prediction models for many properties by combining the information gath-ered from all of these techniques.Data processing plays an important role in the application of spectroscopicprobes for quantitative analysis. In this work, we use a variety of spectra process-ing methods, including discrete wavelet transform (DWT), data fusion, and PLS116regression modeling, to extract meaningful information from the spectra. DWToffers an effective method of spectrum denoising and baseline compression, thatimproves the performance of most calibration models [53, 308]. DWT expressesa spectrum in terms of an orthonormal basis wavelets. Discarding the highest fre-quency wavelets removes high-frequency spectral noise. Discarding the lowestfrequency components reduces the baseline. The midrange frequencies signalingchemical substances of interest are not disturbed.Computational models seek to form classifying representations of target sys-tems. Sometimes, one technique cannot provide enough information to make anaccurate prediction model, so we need to combine multiple data sources to im-prove the accuracy, consistency, and precision of the model. The combination ofdata from different resources has been termed data fusion. This process integratesmultiple data sets that each offer a different measure of some real-world propertyinto a consistent, accurate, and useful representation. Using multiple data sets im-proves the robustness of a model, and leads to more accurate property prediction.Data fusion processes are often categorized as low-level, mid-level or high-level, depending on the processing stage at which fusion takes place [112]. Forlow- and mid-level data fusion, prediction models are constructed after fusing datasets. High-level data fusion fuses completed models, after they have undergonevariable selection and data processing. The expectation is that a fused data set ismore informative and synthetic than the individual original inputs.A PLS regression model predicts values of an observable property y based uponmultivariate measures of its covariance with a multivariate response, X. Standard-ization maximizes the covariance between an X predictor matrix and a vector of yresponses. This approach is especially useful when the number of measured vari-ables in X is much greater then the number of properties collected in y.In this study, we compare NIR, ATR-FTIR, and Raman spectroscopies as ap-plied to a variety of pulp properties: tensile index, burst strength, tear strength,SRE600, light scattering, absorption, wet zero span (WZS), dry zero span (DZS),117and density. We applied and compared low-, mid-, and high-level data fusion tocombine the benefits of using each spectroscopic technique and improve our PLSmodel’s ability to predict these physical and mechanical properties.6.2 Materials and Methods6.2.1 SamplesCanfor Corp. (Vancouver, B.C., Canada) provided 75 northern bleached softwoodkraft (NBSK) pulp sheet samples from the Northwood and Intercontinental PulpMills (Prince George, B.C., Canada). The samples were produced between 2007and 2012, mainly using pulp produced from lodgepole pine (Pinus contorta var.latifolia) and white spruce (Picea glauca) killed by mountain pine beetle (Den-droctonus ponderosae) infestations. Sample preparation consisted of cutting thepulp sheets into roughly 10cm x 15cm cards. The pulp properties to be predictedby spectroscopy with PLS modeling and data fusion were absorption, burst, den-sity, freeness, light scattering, SRE600, tensile index, dry zero span, and wet zerospan.6.2.2 Spectroscopic Instrumentation and MeasurementWe collected Raman spectra with an OLYMPUS BX-51 Raman microscope witha fibre-coupled 785 nm laser source, a motorized microscope stage sample holder,and a CCD detector. Spectra were collected with 30 second integration time and 5co-added accumulations. Each measurement was repeated 5 times.We collected NIR spectra using a Nicolet 6700 FT-IR spectrometer (Thermo-Scientific) equipped with a NIR integrating sphere module and a 5 cm diametersample cup spinner. When operated in the diffuse reflectance mode, this instru-ment illuminates samples with broadband near-infrared radiation from a tungsten-halogen lamp and collects interferograms using an InGaAs detector. The instru-ment’s Fourier transforms span a spectral range from 4000 cm−1 to 9900 cm−1.Each acquisition represents the sum of 64 scans of 10 second integration time. Wecut five circular samples sized to fit the sample cup from different positions in eachpulp sheet, and scanned the top-side of each. Five spectra from each sample were118averaged and used in subsequent analysis.We analyzed five distinct points on each pulp sheet sample using a PerkinElmerFrontier Fourier Transform Infrared / Attenuated Total Reflection (FTIR/ATR)spectrometer. 20 replicate scans were collected and averaged for each point, witha wavelength range of 300 to 4000 cm−1 and resolution of 1 cm−1. ATR-FTIRwas used because of its inherent sensitivity to surface morphology and its abilityto provide information not only about cellulose structures but also about lignin andhemicellulose structures that may be present in samples.6.2.3 Discrete Wavelet TransformDiscrete wavelet transform (DWT) is a well established method of spectral denois-ing and baseline elimination. It provides multi-resolution numerical representationof the original spectrum in waveform convolutions distributed over discrete inter-vals of position. The wavelet transform begins with a mother wavelet, which isdilated and translated to form an orthonormal basis from which the original spec-trum is filtered. Low-pass and high-pass filters remove the high-frequency noiseand the low-frequency background, leaving just the features in the spectrum. AfterDWT ,the spectrum is represented in terms of coefficients c j′,k and d j,k. In thiswork, we use the ’sym5’ basis set with N levels of decomposition. j runs form 1 toN, and k from the lowest to highest frequency for each technique [12, 80, 329].6.2.4 Orthogonal Signal CorrectionOrthogonal signal correction (OSC) minimizes uninformative variation in a signalby finding and removing the features that affect the total variation in a signal andexist in an orthogonal direction to the property of interest [308, 347]. We havediscussed OSC in detail in chapter Template Oriented Genetic AlgorithmTemplate Oriented Genetic Algorithm (TOGA) is a feature selection techniquebased on the genetic algorithm. It reduces the number of features in a signal,and consequently leads to an improvement in PLS prediction accuracy. In thisstudy, we applied TOGA on ATR-FTIR and Raman spectra in order to select the119most highly correlated features to the target of interest for further study [307]. Inchapter 3 and 5, we have explained how TOGA works in detail.6.2.6 Data FusionIn the following section we discuss low-level, mid-level or high-level level datafusion methods. Figure 6.1 shows that different data fusion levels differ by the pro-cessing stage at which the fusion takes place. For low- and mid-level data fusion,prediction models are constructed after fusing preprocessed data sets. The data setsfor low- and mid- data fusion consist of preprocessed data and selected features,respectively. High-level data fusion fuses completed prediction models, after theyhave undergone feature selection and data processing.Low-Level Data FusionWe normalized DWT-NIR, DWT-ATR-FTIR, and DWT-Raman spectra so theywould have a similar scale. As Figure 6.1 represents the normalized spectra areconcatenated together as new spectra for making prediction model of interest. Inlow level of fusion the number of feature in fused spectrum is equal to sum of thefeatures of used individual spectra.Mid-Level Data FusionWe processed and selected important features for each spectra individually. We ap-plied OSC on NIR spectra and TOGA on Raman and ATR-FTIR spectra to selectessential features in predicting the property of interest. As Figure 6.1 represents,in mid level data fusion, we concatenate selected features to make a final spectrumand then build a PLS model to predict the property of interest. This method in-volves a fused spectrum with a much smaller number of features than the one inlow-level data fusion.120                        S1 S2 S3 S1 S2 S3 S1 S2 S3 Preprocessing  Preprocessing  Preprocessing  Preprocessing  Preprocessing  Preprocessing  Preprocessing  Preprocessing  Feature Selection Feature Selection  Feature Selection  PLS PLS  PLS  Preprocessing Feature Selection Feature Selection  Feature Selection  Data Data Data Feature Feature  Feature  Decision Decision  Decision  a) Low Level Data Fusion b) Medium Level Data Fusion c) High Level Data Fusion Integrating Integrating Integrating PLS PLS Figure 6.1: A schematic chart of different levels of data fusion, a)Low level, b)Medium level, and c)High level.121High-Level Data FusionFor high level data fusion which is also known as decision fusion, we integratethe predicted value calculated by the PLS models, made separately for each tech-nique (Figure 6.1). Thus, we constructed PLS prediction models separately foruntreated and treated Raman, NIR, and ATR-FTIR data sets. Then, we calculatedthe RMSECV (root mean square error of cross-validation) and calculated normal-ized weight for each PLS model as follows,WNIR =1/RMSECVNIR1/RMSECVNIR+1/RMSECVAT R+1/RMSECVRaman(6.1)We then scaled the predictions by multiplying each prediction models by its nor-malized weight (6.1). Finally, we fused the weighted predictions by adding thevalues obtained for each technique to calculate the final predicted value. Figure6.2 shows how RMSECVs obtained from each PLS model used to calculate thefinal predicted value for each property of interest.                        NIR ATR Raman DWT  DWT  DWT  PLS 2  PLS 3  OSC TOGA  TOGA  𝑌𝐹𝑖𝑛𝑎𝑙 = 𝑊𝑁𝐼𝑅𝑌𝑁𝐼𝑅 +𝑊𝐴𝑇𝑅𝑌𝐴𝑇𝑅 +𝑊𝑅𝑎𝑚𝑎𝑛𝑌𝑅𝑎𝑚𝑎𝑛  Integrating   PLS 1 YATR & RMSECVATR  YRaman & RMSECVRaman  YNIR  & RMSECVNIR YFinal Figure 6.2: A schematic chart of high level data fusion to calculate the final predicted property of interest.As the chart represents, we calculate W as the weight of each technique to then fuse the result of eachtechniques as the final result.1226.3 ResultsHere, we utilize three different vibrational spectroscopic techniques (ATR-FTIR,Raman, and NIR) in order to asset the effectiveness of each and the combination ofthem in predicting different physical and mechanical properties of pulp. Figure 6.3displays the raw spectra of NIR, ATR, and Raman spectra of a pulp sample. Wesee that both Raman and ATR contain a number of sharp and distinct peaks whileNIR spectra contains mostly broad peaks.The majority of bands in the ATR-FTIR spectra of pulp samples arise fromcellulose (see Fig. 6.3.a). The peak at 3330 cm−1 is assigned to an O – H stretchingband representing the specific O(3)H – O(5) intramolecular hydrogen-bond . Thepeak at 1640 cm−1 corresponds to water molecules. Figure 6.3.a shows C – Hstretching bands in the 3000-2890 cm−1 region, and C – H bending bands in the1500-1300 cm−1 region. The bands at 896 cm−1, 1372 cm−1, 1418 cm−1 are fromthe C – O – C valence vibrations of the glycosidic linkages, C – H deformation incellulose, CH2 scissoring at C(6) in cellulose, and C – H stretching in cellulose,respectively [98, 130].Figure 6.3.b shows the raw Raman spectra for a set of pulp samples. The regionbelow 1500 cm−1 corresponds to the motions of the C – C and C – O framework.We can associate peaks at 1474, 1379, and 1337 cm−1 with the modes that involveconsiderable coupling of methane bending, methylene rocking and wagging, andC – O – H in-plane bending. The peaks in 1090, 1120, and 1150 cm−1 representmodes involving anti-symmetric C-O stretch and ring vibration. We can assign thepeak at 897 cm−1 to H – C – C and H – C – O bending localized at C(6) position inglucose. The region below 600 cm−1 relates predominantly to skeletal, C – C – C,O – C – O, and O – C – C bending vibrations [4, 321].Figure 6.3.c shows representative NIR spectrum of a pulp sample set. We seethat absorption bands associated with many vibrational overtone transitions overlapto form a smooth and relatively undistinguished spectrum. The small peak at 4280cm−1 is attributed to C – H stretching and deformation, while the peak next to itat 4400 cm−1 is assigned to combination bands of O – H and C –– O. The two bigpeaks at 4740 and 5200 cm−1 are associated with O – H and C – H for the first peak(4740 cm−1), and – OH and – C –– O groups for the second peak (5200 cm−1). The123500 1000 1500 2000 2500 3000 3500 4000Frequency (cm−1)Student Version of MATLAB(a) (b)4000 5000 6000 7000 8000 9000 10000Frequency (cm−1)Student Version of MATLAB(c)Figure 6.3: Representative raw ATR, Raman, and NIR spectra in cm−1 for each pulp samples in thepresent datasets. a) Raw ATR-FTIR spectra. b) Raw Raman spectra d) Raw NIR spectra.broad peak around 8000-9000 cm−1 is assigned to second overtone of H2O, C – Hand CH2. The intense wide peak around 6000-7000 cm−1 is ascribed first overtoneof O – H stretching [112, 210, 211].Figure 6.3.a, b, and c are the spectra after DWT treatment for ATR-FTIR, Ra-man, and NIR spectra, respectively. Here, we see an elimination of baseline slopeand offset, together with a significant amplification of reproducible structure overthe full range of the spectrum.Figure 6.5 represents the mid-level fused spectra of three techniques treated124(a)(b)(c)Figure 6.4: Representative DWT treated ATR, Raman, and NIR spectra for pulp samples. a) DWT-treatedATR-FTIR spectra. b) DWT-treated Raman spectra c) DWT-treated NIR spectrawith DWT, which is an assembly of normalized DWT spectra shown at 6.4.a, b,and c. Figure 6.5 shows that all three spectra have the same scale, and can beconsidered as one unit signal.125Figure 6.5: Concatenation of DWT-processed NIR, ATR-FTIR, and Raman data sets for low-level fusion.6.4 DiscussionSince each spectroscopic technique provides some specific information about themorphology and chemical composition of pulp samples, we explored the use oflow-, mid-, and high-level data fusion both to determine the accuracy of each tech-nique in predicting different physical properties of pulp, and the integrate differentprediction models to form a final prediction model of improved accuracy. Ramanspectroscopy offers information probing symmetrical non-polar groups, mid-IRspectroscopy (i.e. ATR-FTIR) yields information pertaining to asymmetric polargroups, and near-IR spectroscopy probes the overtones of the vibrations excitedby mid-IR spectroscopy. Near-IR radiation has the advantage of penetrating muchfarther into sample than mid-IR radiation can, and with appropriate data treatmentsteps, can provide information about the bulk composition of pulp and paper sam-ples. Water has a small Raman cross-section but it disturbs IR spectrum signifi-cantly.1266.4.1 Comparing the Efficiency of Spectroscopic Techniques inPredicting Physical Properties of PulpPredicting Properties with Initial FreenessBoth the properties of a fibre (eg. strength and dimension) and the bonding be-tween fibres have a strong deterministic effect on the properties of paper. Sometests, like wet zero span (WZS) and dry zero span (DZS) measure single fibreproperties while others, like tensile strength, reflect the properties of the fibre net-work. Both single fibre crystallinity and bond strengths between fibres influencefreeness, tensile, and SRE.Several chemical and physical factors affect the strength of paper, includingthe strength and dimensions of single fibres, the build-up of the fibre network,and the bond strength that develops between fibres. Hydrogen bonding betweencellulose chains dominates the process that builds up paper from pulp fibre stock.Other forces, such as dispersion and van der Waals forces between the fibers, alsocontribute to the strength of the paper. The capability of various spectroscopictechniques to probe these types of bonds can determine the technique’s efficiencyfor predicting strength properties.Table 6.1 shows the root mean square error of cross validation (RMSECV) ofPLS models using each of three spectroscopy methods to predict various properties.Poorer sensitivity of a technique for molecular properties that signal to a particularcharacteristic causes a larger error of cross validation. Table 6.1 shows that NIRspectroscopy works better than the other two techniques in predicting of almost allphysical properties of paper samples, since NIR spectroscopy consistently has thelowest error of cross validation.Figure 6.6 shows the fraction of each weighted prediction model that was usedto make the high-level fused prediction model. We calculate the fraction of eachtechnique based on equation 6.1, where the RMESCV for each technique is indi-cated in table 6.1.Figure 6.6 shows that NIR spectroscopy produces a prediction model that is45% more accurate than those made by ATR-FTIR and Raman spectroscopies fortensile strength, burst strength, and scattering. Since these parameters strongly cor-relate with the dimension of the fibres [147], we can conclude that NIR responds127with greater sensitivity to the dimension of fibres than ATR and Raman.Figure 6.6: The calculated weight for each spectroscopic technique, when constructing a high-level datafusion model for different properties. Red: Raman, Green: NIR, and Blue: ATR-FTIRThe wet and dry zero span strength (WZS and DZS) measure the tensile strengthof a single fibre in wet and dry environments. For these two parameters, fibre di-mension has little to no effect. Figure 6.6 shows that Raman and ATR-FTIR spec-troscopies predict the tensile strength of single fibres better than that of paper. Thisshows that both of these spectroscopic techniques are better-suited for looking atsingle fibre properties than bulk material. It also demonstrates that ATR and Ra-man predict properties independent of fibre dimension better than those which varystrongly. NIR spectroscopy still plays a major role in building the fused predictionmodel for WZS and DZS.ATR-FTIR and Raman spectroscopies predict tear strength better than otherstrength properties of paper. If the paper’s fibres are very poorly bonded during atearing test, all the fibres will remain whole following the tear. Thus, interaction128between fibres has high effect on tear strength. Since the fibre properties have haslower influence on tear strength compared to other strength properties of paper,ATR-FTIR and Raman-based prediction models have higher weights in buildingthe fused prediction model, as seen in Figure 6.6.Comparing results in Table 6.1 and Figure 6.6 show that NIR spectroscopypredicts freeness better than Raman, but only slightly better than ATR-FTIR spec-troscopy. Freeness represents the refining degree and draining velocity of paperfibres, and indicates the fibre length of the pulp. Long fibre pulps have higherfreeness compared to short fibre pulps. Freeness depends only on the surface con-ditions and swelling of the fibres. Shorter fibres lead to a rougher surface and areable to interact more with other fibres, which manifests in strong fibre-fibre featuresin the spectra. Since ATR-FTIR spectroscopy is a surface-sensitive technique, wewould expect it to be able to predict freeness well.SRE600 and freeness behave similarly as expected, ATR-FTIR spectroscopysucceeds better in predicting this property than NIR and Raman spectroscopies.SRE600 is defined as the refining energy needed to reach a freeness of 600. Thisdepends on the strength and crystallinity of single fibres [101]. High energy is re-quired to break down fibres and decrease their freeness. Once again, ATR-FTIRspectroscopy makes up the largest part of the fused prediction model for SRE600,although NIR spectroscopy is almost as important.Table 6.1 shows that NIR, ATR-FTIR, and Raman spectroscopies yield rel-atively similar errors of cross validation for water absorption. Water absorptiondepends highly on the crystallinity of the paper. As the crystallinity decreases,the number of hydroxyl groups involved in cellulose-cellulose intarmolecular hy-drogen bonding decreases; they then become free to hydrogen bond with water.ATR-FTIR and Raman spectroscopies both detect the crystallinity disorder of fi-bres [4, 44, 98], and show a RMSEP that compares to the model constructed withNIR spectroscopy.Density is determined by the raw materials used in production, as well as the129Table 6.1: Comparison of the root mean square error of cross validation (RM-SECV) of individual prediction models made with DWT-filtered data.Tensile, burst, tear, SRE, and density have been labeled in different free-nesses as 600 ml, 500 ml, 450 ml, and 300 ml.Parameters NIR-PLS ATR-FTIR-PLS Raman-PLSTensile (initial) 0.2 0.55 0.57Tensile 600 0.32 0.66 0.63Tensile 500 0.38 0.71 0.71Tensile 450 0.43 0.72 0.79Tensile 300 0.57 0.64 0.62Burst (initial) 0.2 0.37 0.39Burst 600 0.31 0.41 0.4Burst 500 0.35 0.64 0.58Burst 450 0.46 0.78 0.73Burst 300 0.38 0.7 0.69Tear (intial) 1.73 2.97 2.77Tear 600 1.19 2.18 2.16Tear 500 1.2 2.11 2.08Tear 450 1.19 1.55 1.54Tear 300 1.12 1.58 1.56SRE 600 12.3 7.92 8.10SRE 500 21.29 12.83 15.08SRE 450 26.77 24.54 26.95SRE 300 33.73 29.56 31.92Density (initial) 0.014743 0.0215 0.0225Density 600 0.01177 0.01653 0.01659Density 500 0.01146 0.01359 0.01385Density 450 0.01158 0.01244 0.0125Density 300 0.0116 0.01187 0.01313Dry zero span 0.46 0.99 0.98Wet zero span 0.5 0.9 0.8Freeness 9.301 9.4 10.5Absorption 0.0085 0.0091 0.0091Scattering 0.81 1.79 1.86130production process itself. The curl and morphology of the fibres (length, roughnessand thickness) affect sheet density [149]. Since the bonding interactions of fibrescan indirectly yield information about fibre morphology, ATR-FTIR and Ramanspectroscopies predict this property with a relatively low prediction error; how-ever, NIR spectroscopy still yields the best prediction model.Predicting Properties with Varying FreenessA paper product that requires a shorter fibre length demands a pulp with a lowerfreeness manufactured with higher refining energy. Refining collapses fibres andmakes them more flexible, also increases their bonding surface area. Smaller andmore flexible fibres better conform to one and other and yield a large fibre-fibrecontact area. Spectroscopic techniques can detect these bonding interactions. Forpulps of decreased freeness and decreased fibre size, bonding interactions in thefibre network become the most important factor in determining different physicaland mechanical properties of paper, instead of single fibre properties. So, for lowerfreeness values and most of properties of interest, all three spectroscopic techniquesshow similar prediction accuracy. However, this is not the case for properties mea-sured for unrefined pulps at initial freeness. In Table 6.1, the freeness at which aproperty is measured is described by the number beside the property. The freenessvalues studied were the initial freeness (unrefined), 600 ml, 500 ml, 435 ml, and300 ml.In this section, we use root mean square error of cross validation (RMSECV)of each technique for each property with different freeness to evaluate the accuracywith which spectra attained by each method to predict different properties of papersheet.Figure 6.6 plots the calculated weights based on equation 6.1 for each of thethree spectroscopic techniques in predicting various properties for different free-ness values. Higher weight means better accuracy in prediction model and higherimpact on final high-level fused prediction model.131Figure 6.6 demonstrates that NIR spectroscopy consistently commands thehighest weight in predicting tensile strength, regardless of freeness. This showsthat NIR spectroscopy responds more sensitively to chemical and morphologicalcomposition, which influence tensile strength. We also see that ATR-FTIR andRaman spectroscopies predicts tensile strength for samples with lower freeness(higher refining energy) better than for unrefined pulp samples. Figure 6.6 confirmsthis by showing that compared to ATR-FTIR and Raman, the impact of NIR spec-troscopy on final fused prediction model in predicting tensile strength decreasesfrom more than 55% for initial freeness to the lower than 35% for freeness of 300ml. Thus, the three spectroscopic techniques differ most in their ability to predicttensile strength for pulps at their initial freeness. This is because lower freenessmeans higher refining energy which forms pulps consisting of smaller and moreflexible fibres. By decreasing the freeness, the fibre dimensions (especially fibrelength) have a smaller effect on paper strength than do fibre bonding and networkstrength. This observation explains the weakness of ATR-FTIR and Raman spec-troscopies in determining properties that depend on fibre dimensions. On the otherhand, ATR-FTIR and Raman spectroscopies predict tensile strength much betterfor lower freeness, demonstrating that these two methods can provide informationthat relates bonding and network strength.We obtained similar results for burst strength. NIR spectroscopy predicts burstfor unrefined pulp and highly refined pulp with freeness of 300 ml, much betterthan the two other techniques.Fibre length and inter-fibre bonding both play an important role in determin-ing the tearing strength. In particular, longer fibres increase tear strength. Theexplanation is straightforward: longer fibres tend to distribute the shear stress oftearing over a greater area, i.e. over more fibres and more bonds, while short fibresconcentrate the shear stress in a smaller area. The strength of chemical bonds inthe fibre for exceed that of the inter-fibre bonds, so most of the tearing events hap-pen between fibres. Shorter fibres reduce the effectiveness of covalent bonds whileincreasing the effectiveness of fibre-fibre bonds, thus lowering the tear strength.132Therefore, while the refining process increases tensile and burst strength, it simul-taneously decreases tear strength. Here, in predicting tear strength we see a reversalof the pattern of accuracy of the three spectroscopic methods. Prediction accuracydecreases from initial to freeness of 600 ml, unlike with tensile and burst strength.However, as with the other properties, decreasing the freeness below 600 ml seesan increase in prediction accuracy using ATR-FTIR and Raman spectroscopies.Density depends a great deal on the morphology and curl of the individual fi-bres. The morphology of fibres in a pulp sheet sample can be estimated by probingthe associated bonding interactions. Knowing the specific freeness value of a sam-ple assures us that the sample has a relatively homogeneous fibre length. Thus, wemitigate the effect of fibre length on final properties of paper by using higher re-fining energies. We can see a similar pattern for density. For initial freeness, a bigdifference emerges between RMSECV observed for NIR spectroscopy with thosefor ATR-FTIR and Raman spectroscopies. This difference decreases dramaticallyfor lower freeness. For a freeness of 450 ml and below, all three techniques yieldsimilar prediction results for density, due to their similar effectiveness in detectingbond interactions between fibres.SRE refers to the refining energy to yield a specific freeness. It depends sensi-tively on the characteristics of single fibres, such as crystallinity. Determining SREfor specific freeness mitigates the effect of fibre dimension. Figure 6.6 shows thatATR-FTIR spectroscopy works as well as NIR spectroscopy for predicting SRE,due to its capacity for determining crystallinity. For the lower freeness, all threetechniques yield similar accuracy in predicting SRE.6.4.2 Fusing Spectroscopic Methods to Build a Final PredictionModelWe have combined data from NIR, ATR-FTIR, and Raman spectroscopies usinglow-, mid-, and high-level data fusion. Since each of these spectroscopic tech-niques provides different specific information about sample properties, we expectthat fusing the data will improve the accuracy of the prediction model. Table 6.2shows that the prediction error of the models using low-level fused data is larger133than the best prediction error which obtained by any of the spectroscopic tech-niques. This does not meet our expectations, and could be due on the high numberof features compared to the number of samples. Partial Least Square (PLS) pre-diction models fail in this type of large data dimension. Decreasing the numberof features using feature reduction or selection techniques help to solve this short-coming of multivariate analysis.In high-level data fusion, we preprocessed the data by DWT (6.4) and madeseparate PLS models for each technique. Table 6.2 shows the accuracy of thehigh-level fused data set compared to the individual data sets. The results implythat high level data fusion resulted in models with similar predictive capability asthe individual spectroscopic model with the best prediction ability, typically thosebased on NIR spectroscopy. However, since each technique is able to predict someof the properties better than the other two, high-level data fusion provide slightlybetter prediction models overall.6.4.3 Fusing Spectroscopic Methods to Build the Final PredictionModel After Reducing the Number of FeaturesIn the previous section, we observed that low-level data fusion failed to improveprediction model accuracy, due to limitations of PLS in forming definitive predic-tion models for data sets with a high number of features compared to the number ofsamples. In this study, the ratio of the number of features to the number of samplesis greater than 65. To decrease the number of features, we used two methodologies:treating each signal separately to reduce the number of its features, and discardingATR-FTIR data in favor of keeping only NIR and Raman signals. We previouslydiscussed the similar performance of ATR-FTIR and Raman in predicting strengthproperties of pulp samples. This similarity motivated us to use only one of thesetechniques, and fuse its data NIR.Figure 6.7 plots the mid-level fused data, consisting of NIR spectra followingOSC-DWT treatment, and Raman spectra following with TOGA treatment. Alltreatments were targeted to tensile strength. As shown previously, OSC enhances134Table 6.2: Comparison of the root mean square error of cross validation (RM-SECV) of prediction models made with individual spectroscopic and datafusion techniques using preprocessed spectra with DWT. Tensile, burst,tear, SRE, and density have been labeled in different freenesses as 600ml, 500 ml, 450 ml, and 300 ml.Parameters NIR ATR Raman H.L. Fusion L.L. FusionTensile (initial) 0.2 0.57 0.61 0.21 0.48Tensile 600 0.38 0.66 0.68 0.38 0.59Tensile 500 0.41 0.71 0.73 0.42 0.65Tensile 450 0.43 0.73 0.76 0.45 0.67Tensile 300 0.51 0.64 0.62 0.52 0.57Burst (initial) 0.22 0.37 0.4 0.22 0.39Burst 600 0.33 0.41 0.41 0.33 0.37Burst 500 0.38 0.64 0.62 0.38 0.66Burst 450 0.44 0.78 0.75 0.45 0.45Burst 300 0.39 0.70 0.69 0.41 0.39Tear (intial) 1.70 2.97 2.81 1.72 1.50Tear 600 1.26 2.18 2.15 1.26 1.21Tear 500 1.24 2.11 2.09 1.26 1.19Tear 450 1.21 1.55 1.56 1.22 1.18Tear 300 1.11 1.58 1.61 1.11 1.10SRE 600 12.333 7.92 8.40 8.42 8.06SRE 500 21.2958 12.83 15.09 12.91 15.54SRE 450 26.8012 24.54 26.95 25.69 26.06SRE 300 33.7233 29.56 31.91 29.83 31.27Density (initial) 0.014 0.022 0.023 0.015 0.13Density 600 0.012 0.016 0.017 0.012 0.11Density 500 0.012 0.015 0.014 0.013 0.013Density 425 0.012 0.013 0.013 0.012 0.010Density 300 0.011 0.011 0.013 0.011 0.010Dry zero span 0.41 0.99 0.97 0.43 0.43Wet zero span 0.50 0.92 0.80 0.56 0.52Freeness 9.33 10.04 10.5 9.41 10.29Absorption 0.0085 0.0093 0.0094 0.0088 0.0095Scattering 0.81 1.77 1.91 0.85 1.77135Figure 6.7: Concatenation of DWT-OSC processed NIR, and TOGA processed Raman data sets for mid-level fusion.the most correlated features for each property, resulting in better prediction model.For Raman spectra, we applied TOGA as an efficient feature selection method. Wedetermined TOGA would be inappropriate for use with the NIR spectra, due to thebroad shape of their signals.The left-hand section of Figure 6.7 illustrates NIR spectra after treatment withOSC-DWT for initial tensile as target. Comparing the OSC-DWT-NIR featureshere with DWT-NIR in Figure 6.7 shows that some parts of NIR spectra has beenenhanced while other parts have been suppressed. OSC-DWT-NIR spectra man-ifest a large variation in the region 4300-5800 cm−1, with an isosbestic point at4700 and 4900 cm−1.Figure 6.4.b shows DWT-treated Raman spectra. The DWT spectra do notshow any feature that stands out as a more important univariate feature correlatedwith a specific property of pulp. On the other hand, the right-hand parts of theFigure 6.7 represents the TOGA-reconstructed spectra of Raman, using initial ten-sile as target. Figure 6.7 demonstrates how TOGA can isolate the set of distinctfeatures that relate most to a selected property, in this case initial tensile strength.Table 6.3 lists the RMESCV of prediction models built using NIR and Raman136Table 6.3: Comparison of prediction models made with individual spectro-scopic and data fusion techniques. In this table, Raman spectra havebeen treated with TOGA and NIR spectra with OSC, as chemometricfeature selection methods. Tensile, burst, tear, SRE, and density havebeen labeled in different freenesses as 600 ml, 500 ml, 450 ml, and 300ml.Parameters OSC-DWT-NIR TOGA-Raman M.L. Fusion H.L. FusionTensile (initial) 0.17 0.39 0.14 0.18Tensile 600 0.37 0.49 0.31 0.37Tensile 500 0.38 0.57 0.37 0.38Tensile 450 0.4 0.61 0.41 0.40Tensile 300 0.44 0.52 0.39 0.47Burst (initial) 0.22 0.24 0.19 0.22Burst 600 0.31 0.31 0.27 0.31Burst 500 0.35 0.41 0.33 0.36Burst 450 0.43 0.52 0.41 0.44Burst 300 0.37 0.49 0.34 0.37Tear (intial) 1.71 1.92 1.73 1.71Tear 600 1.24 1.37 1.22 1.24Tear 500 1.23 1.41 1.2 1.26Tear 450 1.21 1.09 1.21 1.22Tear 300 1.07 1.12 1.05 1.09SRE 600 12.21 5.73 5.92 6.80SRE 500 19.87 11.66 11.57 13.74SRE 450 23.87 22.35 20.86 22.32SRE 300 32.11 27.15 25.99 28.42Density (initial) 0.013 0.017 0.012 0.013Density 600 0.010 0.013 0.009 0.011Density 500 0.009 0.01 0.009 0.009Density 450 0.01 0.008 0.007 0.008Density 300 0.01 0.007 0.007 0.008Dry zero span 0.40 0.76 0.40 0.40Wet zero span 0.50 0.64 0.49 0.50Freeness 9.42 9.78 9.33 9.48Absorption 0.008 0.008 0.008 0.008Scattering 0.81 1.22 0.79 0.81137Figure 6.8: The calculated weight for OSC-DWT-NIR and TOGA-Raman spectra in building a high-leveldata fusion model for different properties. Tensile, burst, tear, SRE, and density have been labeled indifferent freenesses as 600 ml, 500 ml, 450 ml, and 300 ml. Red: Raman, and Green: NIR.spectra after OSC and TOGA signal processing techniques, respectively. It showsthat the processing steps improved the models built by both techniques. TOGA im-proved the efficiency of prediction model dramatically, by decreasing the numberof features in the Raman spectra. The values in Table 6.3 suggest that NIR is still abetter-suited technique for predicting almost all of the properties. However, as Fig-ure 6.8 illustrates, the contribution of Raman in making the final fused model sig-nificantly increased after applying TOGA compared to using simply DWT-Raman.Table 6.3 shows that mid-level fusion slightly improves the prediction modelover the individual methods.1386.5 ConclusionThe present work uses data fusion to determine whether combining NIR, ATR-FTIR, and Raman spectral information provides prediction models with improvedefficiency in predicting physical and mechanical paper properties, compared tomodels built using only the individual techniques. We have compared the effec-tiveness of these three spectroscopic techniques in predicting pulp properties byapplying the root mean square error of prediction to evaluate PLS models. Wehave found that NIR spectroscopy performs most favorably in predicting mechani-cal properties. However, NIR spectra offer overlapping spectra with limited infor-mation available, so it is not an effective technique for spectroscopic troubleshoot-ing. ATR-FTIR and Raman both showed similar performance in predicting pulpproperties with similar accuracy in prediction. Raman spectroscopy is particularlywell-suited to pulp and paper analysis. It extracts chemically specific signaturesfrom aqueous samples with little or no sample preparation, while NIR and ATR-FTIR techniques both show an overwhelming broad peak for water. We also foundthat Raman and ATR-FTIR spectroscopies predict properties with higher beatingand lower freeness more accurately than NIR. This observation indicates the abil-ity of Raman spectroscopy to detect hydrogen bonding interactions present in thesample. Since refining collapses fibres, they become more flexible and their bond-ing surface area is increased after refining. By increasing the refining energy toyield smaller fibres, they interact with one another to a greater degree. Detectinghydrogen bond interactions becomes more important in classifying such samples,which can be seen by the improved analytical performance Raman and ATR-FTIRspectroscopies.139Chapter 7Live, Three-DimensionalDynamics of Nanoscale ParticleInternalization, DetectedChemically and Morphologicallyin Human Cells7.1 IntroductionRapid advancements in microscopy have allowed scientists to resolve subcellularcompartmentalization and ultimately individual molecules [3, 25, 64, 128]. Cur-rently, a variety of optical imaging techniques are available for the study of biolog-ical systems.Fluorescence microscopy represents a broad category of optical imaging tech-niques. The development of fluorescent proteins and fluorophore-conjugated anti-bodies has provided biologists with the tools to selectively label individual moleculesand track their positions within a cell. While illumination methods and signalcollection vary between the different techniques, in all cases, a fluorophore is in-troduced to the cell, which absorbs an incident photon (or photons) at an excita-140tion wavelength and emits a photon at a different wavelength following an internalchemical relaxation [152, 322, 336].The simplest form of fluorescence microscopy is Laser Scanning FluorescenceMicroscopy (LSFM) [116, 240]. In this method, a monochromatic excitation laser(usually in the UV-visible region of the spectrum) is rapidly scanned over the sam-ple, exciting fluorophores in the beam path. Upon relaxation, the fluorophoresemit photons at a longer wavelength than the excitation laser, which are redirectedto a photodetector using a dichroic mirror. In most instrumental setups, confo-cal pinholes remove fluorescence originating from above or below the focal plane.Confocal laser scanning microscopes are easy to implement, have fast image acqui-sition times, are capable of three-dimensional image reconstruction (with the useof a translational stage), and offer diffraction-limited spatial resolution. However,using a short wavelength excitation laser increases potential photo-damage to thesample or fluorophore itself [116, 267].Multiple-Photon Excitation Fluorescence Microscopy (MPEF) operates simi-larly to confocal LSFM with a few notable differences. Instead of using a singlehigh-energy excitation photon, fluorophores absorb two incident photons carryinghalf as much energy each; thus, the same excitation effect is produced provided thephotons reach the fluorophore within approximately 1 fs of each other. This offerstwo distinct advantages over confocal LSFM. First, the use of a longer wavelengthexcitation laser (typically in the near-infrared spectrum) reduces spontaneous scat-tering and absorption of incident photons, thereby increasing the depth of field thatcan be imaged. Second, because of the femtosecond timescale of the excitationevent, fluorescence is restricted to the areas of high photon density at the sharpfocal point of the laser. Thus, all detected fluorescence comes from the focus, andconfocal pinholes are not required [228, 286, 300, 346, 363].The aforementioned fluorescence methods create two-dimensional images byscanning a single point excitation source over the field of view, which typicallyrequires significant time. Light sheet microscopy (LSM), on the other hand, illu-minates a full plane of the sample and acquires all of the resultant fluorescenceby means of an objective aligned perpendicular to the illumination. LSM pro-duces a high signal-to-noise ratio and is capable of creating three-dimensionalreconstructions, with proper manipulations to the sample holder or illumination141source. In addition to conventional fluorescence limitations, LSM requires uniquesample preparation that is not suitable for all samples, and generates very largedatasets that may result in computational processing problems for larger projects[253, 256, 262, 325].The spatial resolution of conventional fluorescence methods, including as thoselisted above, is constrained by the diffraction limit of light (approximately half thewavelength of the detected light), arising from the point spread function (PSF) ofa point source of light. Recent experimental designs have succeeded in breakingthis limit, and are termed super-resolution microscopy techniques [133]. The mainsuper-resolution techniques in use are stochastic reconstruction microscopy meth-ods, Stimulated Emission Depletion Microscopy (STED) [121], and Structured Il-lumination Microscopy (SIM) [39, 129].Stochastic Optical Reconstruction Microscopy (STORM), Photoactivated Lo-calization Microscopy (PALM), and Fluorescence Photoactivation Localization Mi-croscopy (FPALM) are super-resolution techniques collectively referred to as Stochas-tic Reconstruction Microscopy. Two fluorophores that cannot be spatially resolvedby conventional methods become resolvable temporally by applying stochasticreconstruction microscopy methods. Progress in engineering and modeling z-dependent PSFs have also realized three-dimensional reconstructions at super res-olution [227, 361]. While offering some of the highest spatial resolution available,STORM/PALM/FPALM are limited by the efficiency of sample labeling, the re-liability of photoactivatable tags, and are difficult to extend to temporal studies[10, 227, 361].The fluorescence signal detected in STED is restricted to a much tighter focusat the center of the excitation beam, and careful rastering of the sample throughthis focus allows for spatial resolution beyond the diffraction limit. As such, widefields of view require longer acquisition times, and the method is not suitable forreal-time applications. Higher laser intensities used in STED may also pose greaterconcern for photobleaching of fluorescent labels [235, 312].SIM uses a diffraction grating to generate wide-field illumination with a knownsinusoidal intensity pattern. The recorded fluorescence intensity is then the productof the excitation intensity with the density of fluorophores in that region. Varyingthe orientation of the patterned illumination modulates the fluorescence accord-142ingly. Applying a Fourier transform to the data set provides spatial information atup to twice the resolution of the detected light, which can then be used to recon-struct the sample image at this enhanced resolution. While SIM benefits from lowlaser power and wide-field imaging to produce super-resolution reconstructions,the underlying principle of the method means that spatial resolution cannot surpasstwice the diffraction limit, effectively imposing a new diffraction-limited resolu-tion. The biggest limitation of all fluorescence methodologies is their inherentneed to label the sample. With these limitations in mind, there is a great interest indeveloping purely optical, label-free methods for visualizing cells, while retainingthe advantages of fluorescence methods [168, 235].Many different microscopy techniques are available which do not require label-ing. Bright-field microscopy is the simplest and most prevalent technique, relyingon absorption of white light by the sample. Bright-field is an absorbance basedmicroscopy; absorbing structures in the sample reduce the intensity of the lightreaching the objective, resulting in dark features on a bright background. The aque-ous interior of a cell has a very small absorbance cross-section for white light, andthus generates insufficient contrast to be visualized in bright-field. To counteractthis, cells are often stained with cytotoxic contrasting agents, making bright-fieldunsuitable for in vivo studies. The incoherent white light source also imposes adiffraction limited spatial resolution and prevents three-dimensional optical sec-tioning. As such, bright-field is typically used to locate features of interest to befurther analyzed by more sophisticated techniques [277, 306, 362].Recognizing the limitations of absorbance-based microscopy, dark-field mi-croscopy produces an image from light scattered by the sample, rather than ab-sorbed. Dark-field produces a wide-field image with diffraction limited resolutionand is well suited to label-free studies of live cells, as their biochemical complexityproduces large scattering cross-sections. However, the high illumination intensityrequired to generate sufficient contrast in dark-field may limit its use in photo-sensitive samples. Difficulties in optical sectioning also make three-dimensionalreconstruction difficult to perform [331, 334, 335].Relying on neither absorption nor scattering, differential interference contrast(DIC) microscopy maps changes in a sample’s refractive index to create an image.The resulting image produced in a DIC experiment is a shadowed trace of refrac-143tive index boundaries with sharp contrast, often difficult to detect by conventionalbright-field. Like dark-field, DIC offers wide-field images suitable for live cell,label-free experiments. Work on data processing of DIC images has shown three-dimensional reconstructions to be possible. However, this technique is not suitablefor thick samples [11, 14, 61].The label-free nature of the aforementioned optical microscopy techniqueshave so far been presented as advantageous over fluorescence techniques. How-ever, in a fluorescence experiment, the observed signal arises from fluorophoresspecific for a defined molecular target. Eliminating external label requirementscomes at the cost of molecular specificity. Raman microspectroscopy, or Ramanmapping, acquires chemical information about the sample while remaining suit-able for label-free, in vivo experiments. Further processing of the data, via peakintegration or multivariate analysis, can transform the spectral intensity data intoa spatial distribution map for a chemical marker. While chemically specific three-dimensional reconstructions at diffraction-limited spatial resolution are possiblewith Raman mapping, the method is limited by long data collection times. A sin-gle spectrum may only take a second to acquire, but extending this to a wide fieldof view over two or three dimensions at nanoscale step-sizes rapidly increases theruntime of an experiment. Thus, Raman mapping is unsuitable for real-time imag-ing of cells. However, it may be used to monitor spectral changes at a single focusover a prolonged period of time [1, 187, 278].Another limitation of Raman microspectroscopy is the weak signal intensitygenerated by spontaneous Raman scattering. Multiple variations of Raman spec-troscopy have been developed that exploit resonant frequencies or use multiplelasers to enhance the signal, such as coherent anti-Stokes Raman spectroscopy(CARS) [64, 122, 244, 365], resonance Raman spectroscopy [16, 292], tip-enhancedRaman spectroscopy [20, 219, 271, 354], and surface-enhanced Raman spectroscopy(SERS). SERS has proven to be a popular method among cell biologists, as label-ing samples with noble metal nanoparticles can increase Raman signal intensity bymany orders of magnitude. Introducing AuNPs to the sample provides more de-tailed information when using surface enhanced Raman spectroscopy (SERS). InSERS, coupling of the surface plasmon resonance of a nanoparticle produces moreintense and sensitive signals in the area surrounding the particle. This enhances the144detail generated in a Raman mapping experiment, and can be used to measure thenumber of AuNPs internalized by the cell [66, 141, 160, 221, 296, 311].Raman spectroscopy is a powerful method for the label-free imaging of livecells, but is limited by long acquisition times. In a Raman mapping experiment, itis important to validate that the system being studied does not change dramaticallydue to the long timescale or laser excitation energy.Interferometric scattering microscopy (iSCAT) is a complementary imagingtechnique that suits this purpose well. Briefly, iSCAT collects back-scattered lightfrom the sample, and compares its phasing information to a reference field to createan image, similar to conventional dark-field microscopy [57, 229]. The result is awide-field view of a label-free live sample capable of video frame rates in realtime. In this article, we discuss the uptake of AuNPs in a HeLa cell model and itssubsequent imaging by Raman mapping and iSCAT.7.2 Experimental Design7.2.1 iSCAT-Raman InstrumentationFigure 7.1 shows a schematic diagram of the combined interferometric scattering(iSCAT)-Raman microscope used in the present research. This instrument rep-resents a novel combination of high-resolution wide-field real-time optical imag-ing and label-free, non-destructive confocal Raman microscopy. Together, thesetechniques readily provide chemical and morphological information about sampleswith minimal preparation needed. To our knowledge, this instrumental approach isunique [57]. Confocal Raman spectroscopy has been widely reported in the liter-ature [45, 55, 58, 84, 105, 158, 246]; details about iSCAT’s mechanism of actionare discussed in a later section.iSCAT PathAfter exiting the green 532 nm laser (LaserQuantum ltd), the iSCAT illuminationbeam is enlarged and rastered over a rectangular area by the two perpendicularacousto-optic beam deflectors (Gooch & Housego plc). The rastered beam is fur-ther expanded and directed into the infinity-corrected high-NA oil immersion ob-145jective (Olympus Corp.). Backscattered light from the sample is returned alongthe same path. It is segregated from the Raman path by long-pass filter, and sepa-rated from the incident light beam by a 50:50 beam-splitter cube, which directs thebackscattered light into a CMOS camera (Point Grey Research, Inc.) with a 100 x100 µm2 field of view.Confocal Raman Path633 nm red light leaving the HeNe laser (Thorlabs, Inc.) is expanded and colli-mated through a spatial filter and directed to fill the objective. Backscattered lightreturning from the sample follows the same path, passing through the iSCAT long-pass filter unperturbed. At the second longpass filter, Rayleigh-scattered light isremoved from the beam, and the remaining Stokes-scattered light passes throughthe confocal spatial filter and into the spectrograph (Princeton Instruments). Aclosed-loop high precision translation stage (Piezosystem Jena GmbH) rasters thesample with respect to a fixed objective focus.Data CollectionThough the iSCAT and Raman branches are coaxial, the overwhelming power ofiSCAT illumination means that iSCAT and Raman data cannot be collected simul-taneously. However, switching between iSCAT and Raman modes is facile.iSCAT imaging data can be collected as still images or video; video is prefer-able for recording real-time dynamics within a sample. Both collection modes canbe used during piezo scans, where the automated piezo stage rasters the samplealong a user-defined path. The most useful type of iSCAT scan is along the opticalz axis, i.e. a depth scan. Collecting a series of iSCAT images (or frames in a video)at different focal depths can illuminate the sample three-dimensional resolution.Collecting spectral data with the confocal Raman branch is inherently a point-by-point process. It is often useful to use an iSCAT image to determine sampleregions that suggest structure of chemical of interest; the Raman probe can then bedirected to those regions using the piezo stage. In order to build a Raman map thatwill be analogous to an iSCAT image, the sample can be rastered in two dimensions(xy) along a user-defined path, with a spectrum collected at each point.146Figure 7.1: Schematic diagram of the combined iSCAT-Raman microscope,including lasers operating at 532 nm (iSCAT) and 633 nm (Raman),acousto-optic beam deflectors (AOBD), long-pass filters (l p f ), CMOScameras for iSCAT and Bright-field, confocal apparatus, and spectro-graph. Note that the bright-field system is peripheral and was not usedin the present research.7.2.2 Sample PreparationWe cultured HeLa cell line in Dulbeccos modied Eagle’s medium (DMEM) con-taining 10% fetal bovine serum. Thin coverslips were soaked with 70% ethanoland dried in a sterile air flow in the biosafety cabinet. They were next treated with0.1% sterile gelatin, and dried again. Treated glass was placed in petri dish filledwith medium and serum. The cells were loaded and incubated overnight at 37 ◦Cin 5% CO2.Preparation of slides with normal human lung fibroblast HFL1 cells (UBCChemistry Cell Line Collection # 846) was similar to the preparation of HeLacells, except we used Ham’s F-12K (Kaighn’s) Medium with 2 mM L-glutamineand 10% fetal bovine serum. To prepare slides containing both HeLa and normalfibroblast cells, we followed the same steps as HeLa cell preparation, culturing thecells in DMEM, containing 10% fetal bovine serum.1477.2.3 Image ProcessingWavelet transform (WT) plays an increasingly important role in signal and im-age processing. Wavelet transform has several useful applications such denois-ing, resolution enhancement, reconstruction of missing regions, image compres-sion, edge detection, feature extraction, and 3D object processing [12, 72]. Inthis study, we apply wavelet transform to denoising, resolution enhancement, andthree-dimensional image processing using iSCAT images.Two-Dimensional Wavelet TransformDiscrete wavelet transform (DWT) decomposes a signal into two functions of scaleand time, providing a multi-resolution time-scale dataset using a basic waveletfunction W (t) (equation 7.1).Wm,k(t) =1√aW (1a(t−b)) (7.1)Where m is the scale, k is the location of an image or a signal that wavelet isapplied to, a is equal to 2m, and b is equal to k∗2m. In this study, we used Gaussianwavelet, as shown in Figure 7.2-a.Wavelet transform uses such a function as equation 7.1 to decompose an im-age into a set of different frequencies. Figure 7.2 presents the principle of imagewavelet decomposition with two decomposition steps. To do so, a high-pass andits complementary low-pass filters are applied to image columns, followed by thesame process on image rows. Each column or row can be considered as a onedimensional signal with N elements, as per equation 7.2:[ImageN×N]=R(1)R(2)...R(N)= [C(1) C(2) . . . C(N)] (7.2)For the first level of the decomposition, high- and low-pass filters can be ap-plied to produce two different frequency signals (high and low frequency) with halfthe number of elements (N/2). We apply the same procedure on rows producing148four different images: low-low (LL), low-high (LH), high-high (HH), and high-low(HL) subimages. For the next level of decomposition, the low-low subimage canbe further decomposed by high- and low-pass filters [48, 68, 72, 143].Image DenoisingWavelet Transform decomposes an image into a set of low and high frequencies,and after applying threshold coefficients to filter out unwanted noise, the denoisedimage can be reconstructed. Threshold coefficients determine which frequenciesare considered to be noise noise and which contain essential image data. Noise iniSCAT images consists of random signals which are unrelated to the object of inter-est. In this study, the objects of interest are mostly represented in the low-frequencydomain, while noise is dominant with the high-frequency domain. Therefore, usinglow-pass filter, can remove high-frequency noise and reconstruct an image consist-ing of a determined range of “useful” frequencies. We removed high-frequencynoise by setting a threshold for wavelet transform. Based on what informationis useful and what is not, we can keep or remove any specific frequency (with apredefined resolution).Wavelet Transform and Image Quality AssessmentWe measure the quality of an image to determine the extent of its blurriness. To dothis, we use wavelet transform multi-resolution analysis, requiring no informationfrom the original image.Tong et al. report that blurriness mainly impacts the type and sharpness of anedge. They propose that blurred images contain different types of edge, and thesharpness of an edge represents the extent of the overall blurriness [314]. Edgesare high-frequency features in the image [69, 314]; therefore, wavelet transformseems to offer a good means to isolate and represent blurred edges.In general, four categories of edge are available in an image:• Dirac structures, which for blurred edges become Roof structures.• Step structures, which can be classified as either A- or G-step structures. ForA-step structures, the change of intensity is very sharp, while it is gradual149(a)(b)Figure 7.2: (a): Gaussian wavelet used for 2-D Discrete Wavelet Transform (DWT). (b): iSCAT image ofHela and fibroblast cells, after applying 2-D DWT to discrete image in different frequency levels. Thefirst-level low-low subimage (LL1) has been processed with 2-D DWT again to produce second-levelsubimages.150for G-step structures.Figure 7.3 shows the four types of edges. For G-step structures and roof struc-tures, a parameter a is defined to determine the sharpness of the edge. A larger aindicates a sharper edge.(a) (b)(c) (d)Figure 7.3: Graphical representations of edge structures. (a): Dirac structure (b): Roof structure (c):A-step structure (d): G-step structureOut of focus features cause blur by converting Dirac-structures to Roof-structures,and A-step structures to G-step structures. G-step and Roof structures do notchange when blurring occurs; however, their sharpness does decrease (smaller a).Calculating the percentage of edges that are G-step and Roof structures representsan estimation of the blur extent in an image. To do so, we first decomposed theimage with Gaussian wavelet transform to four levels (Fig. 7.2-a). Afterward, wecalculated the edge map Ei according to equation 7.3:Ei =√LH2i +HL2i +HH2i (7.3)where i (scale) is iterated from 1 to 4 [314].151We map Ei for all the four scales of Gaussian wavelet, and then partition themap into windows of fixed size. Based on the level, the size of the window typicallyranges between 2× 2 for the highest scale (i = 4). and 16× 16 for lowest scale(i = 1). Local maxima in the edge map Ei represent the intensity of edges. We seta threshold Thi against which we can compare the intensity of these maxima.Comparing the results of each decomposition scale (i.e. each resolution) pro-vides us information about the extent of the blurriness in an image. We considermaxima in Ei to be edges if their intensity is higher than the selected threshold in atleast one scale. To select an appropriate threshold, Kerouh and Serir have proposedequation 7.4:Thi =2i−1M ∗N ∗M∑k=1N∑l=1Ei(k, l) (7.4)where M×N is the image size [157].Comparing the intensity of maxima in different levels indicate the type of theedge as follows. If Emax-1 < Emax-2 < Emax-3, then the edge has a G-step or Roofstructure, while if Emax-2 > Emax-1 and Emax-2 > Emax-3, the edge has a Roof struc-ture. To consider a G-step or Roof structure to be blur, the first level calculatedEmax-1 should be greater than the selected threshold. However, to have a rough ideaof the blurriness of an image, we use the ratio of the sum of the Dirac and A-stepstructures over all the detected edges to represent the extent of sharpness of theimage [314].The edge map Ei is positive for possible edges, and is equal to zero for non-edgeareas. Note that the edge pixel intensity depends on the level of decomposition.Thus, for each decomposition level, a specific threshold should be set.Blur ReductionBased on the image reconstruction method, blurred edges can be sharpened orremoved outright. Removing all blurred edges provides two-dimensional imagescontaining only in-focus features; these can be compiled into an informative three-dimensional volume. However, to obtain the best representation of a sample in asingle two-dimensional image, it is preferable to retain the out-of-focus informa-tion contained in blurred features; that information is best accessed by sharpen-152ing blurred edges. The primary interest of this study is the construction of three-dimensional volumes of our samples; therefore, we removed blurred features fromour images, retaining only sharp, in-focus information.Single Particle TrackingSingle-particle tracking techniques have allowed us to observe the structure anddynamics of an individual nanoscale objects. iSCAT does not directly image goldnanoparticles in cells, but rather images the convolution of the nanoparticle with itspoint spread function. The point spread function degrades the quality of an image.Localizing the center of mass of a single particle can improve the optical resolu-tion of the system. This can be accomplished by computationally deconvolutingthe point spread function from the observed image. One approach is to fit the pointspread function of a particle to a two-dimensional Gaussian function, which facil-itates deconvolution by allowing the point spread function to be mathematicallymodeled and removed for any particle in an image. In the absence of external fluc-tuations, the signal to noise ratio limits the speed and precision with which we cantrack the point spread function of a single molecule [8, 293, 294].To localize a particle, we follow two steps: we locate the brightest pixels in theimage, and then we calculate their centroid-weighted position.To obtain an initial estimate of the location of a particle, we start with thebrightest pixel of each particle. To do so, we consider pixels in an area with aradius smaller than the distance between two separate particles and larger than theradius of a particle itself. In that radius, we select the brightest 60% of pixels torepresent the particle. The plot of particle pixel intensity versus the pixel numberalong one dimension (x or y) falls in a Gaussian distribution. However, in mostcases, the brightest pixel is not at the centroid of this Gaussian plot. We use theone-dimensional Gaussian distribution and its centroid to localize the x-coordinateof each particle, and repeat thus procedure to localize the particle’s y-coordinate.By repeating this procedure in neighboring frames, collected at 45 frames persecond, we can track gold nanoparticle motion in real time. For some points, how-ever, no coordinate information can be determined, due to a particle’s diffusion intoand out of the focal plane of the image.1537.3 ResultsWe have investigated the performance of a new technique that combines interfer-ometry scattering microscopy (iSCAT) with confocal Raman spectroscopy [57].Three different features are of the interest to this study:1. The three-dimensional structure of different cells illuminates by iSCAT.2. The dynamics of the interaction between a live cell and a gold nanoparticle,in three steps: while entering the cell, while in the cell membrane, and whilein the cytoplasm3. The chemistry of the regions of the cell in which gold nanoparticles aggre-gateThe combination of iSCAT and Raman data allows us to address these points.Each is covered in the following sub-sections.7.3.1 Nanoparticle Dynamics Before, During, and AfterInternalization by Live CellsWe evaluated the utility of the iSCAT-Raman approach by studying the dynamicsand interactions of gold nanoparticles with live cells. For this reason, we usediSCAT to observe the internalization of gold nanoparticles (AuNPs) by live cells,and the motion of the nanoparticles within the cells.We investigated the interaction of 30 nm gold nanoparticles with HeLa cellsas a model cancer cell. Figure 7.4 contains 12 frames collected over a timescaleof one minute, illustrating the two gold nanoparticles entering the membrane ofthe HeLa cell. Figure 7.4 is consistent with the idea that internalization of a goldnanoparticle is time-dependent.As our goal was to test the capability of iSCAT in observing and capturing time-dependent events, the details of nanoparticle endocytosis are beyond the scope ofthis study. Such information has been previously reported.[23, 24, 190, 358]To evaluate the feasibility of iSCAT in detecting gold nanoparticles associatedwith cells, we recorded an iSCAT video over 2 hours after adding 30 nm gold154Figure 7.4: Successive frames showing the interaction of two gold nanoparticles with the membrane ofHeLa cell before internalization. Only a small part of the cell membrane is shown. The temporalinterval of each frame was 10 s. The scalebar is 200 nm.nanoparticles to a sample containing live HeLa cells. This video provided infor-mation on the dynamics of gold nanoparticle motion inside HeLa cells.The large, top image in Figure 7.5 shows two HeLa cells attached to one an-other. Internalized 30 nm gold nanoparticles are easily detected and identified in inthe membrane of the cells. Only a few of the gold nanoparticles are observable inthe cytoplasm.155Figure 7.5: Gold nanoparticle dynamics in HeLa cell membranes. Top: Over-all image of nanoparticles and cells. Bottom: Frames showing the mo-tion of nanoparticles inside the cell membrane; expansion of the illus-trated area in the top image. The temporal interval of each frame is 7s.The scalebar for top and bottom images is 10 µm and 200 nm, respec-tively.156The lower, small images in Figure 7.5 show a time series of the dynamicsof gold nanoparticles in a HeLa cell. The frames show that particles are movingindependently of each other. However, they only move within a small part of thecell during the time we recorded the video.7.3.2 Chemistry of Nanoparticle-Dense Areas Within CellsThe cell contains of four major chemical groups: proteins, nucleic acids, lipids,and carbohydrates. In this study, Raman spectroscopy observes the chemistry ofeach part of the cell. Biological molecules have Raman shifts in the range of 600-3000 cm−1. Figure 7.6 shows a surface-enhanced Raman signal of a HeLa cell; thecomplex spectrum illustrates the different chemical groups in the cell.The dominant peaks related to carbohydrates in the HeLa cell are 1025, 1081,and 1152 cm−1, corresponding to the coupling of the C – O stretching and C – O – Hdeformation modes assigned to glycogen. Smaller peaks at 1070, and 1656 cm−1are attributed to collagen.The peak at 1096 cm−1 is assigned to the phosphodioxy group. The peak at788 cm−1 is assigned to the phosphodiester bond in DNA, and the peak at 813cm−1 is assigned to phosphodiester in RNA identify the chemistry of nucleus. Thepeak located at 782 cm−1 is assigned to thymine, cytosine, and uracil. The peakat 1578 cm−1 shows the existence of guanine and adenine in the nucleus. Smallerpeaks at 728, 1374, 1421, and 1486 cm−1 are also correspond to the nucleus.The Raman peaks at 1301 cm−1 and 1499 cm−1 correspond to C – H vibrationsin the hydrocarbon chain in lipids. The peak at 1660 cm−1, assigned to C –– Cstretching, also represents lipid components. Other lipid peaks are located 719,1743, and 2852-2888 cm−1, attribute respectively to choline, C –– O stretching, andsymmetric CH2 stretching.Proteins in the cell can be identified by dominant peaks at 1666-1670 cm−1,which are assigned to Amide I, and 1200-1300 cm−1, which assigned to Amide III.Raman peaks related to phenyl groups in amino acids also can be used to recognizeprotein in a cell. Some of the main phenyl group peaks are 1005 and 622 cm−1,assigned to phenylanaline, 877 and 854 cm−1, assigned to tyrosine, and 760 cm−1assigned to tryptophan.157Figure 7.6: Surface enhanced Raman Spectra of different parts of a HeLa cell. Exposure time: 0.5 s.Laser wavelength: 633 nm. black: lipid, blue: nucleic acid, red: protein, and green: carbohydrate.A major component in cellular membrane is phosphatidylcholine, which ex-hibits a Raman peak at 719 cm−1. Using these peak assignments, different cellregions can be segregated by building a map of Raman spectra collected at pointsacross the cell, and those identified regions can be correlated with features observ-able in iSCAT images.7.4 DiscussionWe present a new interferometric light microscopy system combined with Ra-man spectroscopy. This wide field imaging system has successfully demonstratedthree dimensional localization and spectroscopic identification of domains in goldnanoparticle-targeted human cells. Spatial and spectral information of this typecould offer utility for such biomedical applications, as disease detection and treat-ment in complex biological environments. The present system can be used forreal-time three-dimensional nanoparticle tracking in which the surface-enhancing158Figure 7.7: iSCAT image of a fibroblast cell. Left: Raw iSCAT image. Right: Deconvolved iSCAT imageusing DWT. The scale bar is 10 µm.ability of nanoparticles act as spectral sensors.7.4.1 Three-Dimensional Reconstruction of CellsiSCAT microscopy provides a means to image living cells in real time withoutthe need for any labels, offering an opportunity to study biological mechanismsand dynamics. Moreover, a combination of a series of images taken along theoptical z axis can yield a three-dimensional structure of the specimen. Each two-dimensional (xy) iSCAT image of a sample contains information from both thestructure of the in-focus part, and the intensities from the out-of-focus light origi-nating from neighboring image planes.To remove these out-of-focus features from each xy plane, we used deconvolution-based image processing. In the following section, we discuss the efficacy of iSCATin capturing three-dimensional volume image of a cell, in studying the dynamic ofgold nanoparticles interacting with live cells, and in providing structural-chemicalinformation about any chosen part of a cell.The construction of a three-dimensional image requires many focal plane stacksthat are clear of out-of-focus haze. However, as Figure 7.7 illustrates, the raw iS-CAT image contains a persistent non-uniform background, noise, and blurred fea-tures. Fig. 7.7 also shows that the contrast of the raw iSCAT image is relativelylow. The impaired quality of the raw iSCAT image is due to the interaction of159 10 μm Figure 7.8: HeLa cells affixed to the cover slip, after removing background and noise. This image hasbeen obtained by calculating the sum of 15 different z planes (0−10 µm), showing HeLa cell structurein the x, y, and z dimensions.the backscattered light with many optical components. The beam passes severallenses, mirrors, and beam-splitters before reaching the detector, and each compo-nent has a small detrimental effect on beam quality. These negative contributionsare constant for a given incident light power; thus, they can be accounted for andremoved. To do so, we take an iSCAT image of a plain coverslip, and use it as arepresentative background signal, which can be removed from other images.Other factors which deteriorate the image quality are noise and blurred fea-tures. Each focal plane is contaminated with out-of-focus features from neigh-boring planes, causing blurriness. In this study, we use wavelet transform to firstdenoise the image by applying a threshold to wavelet coefficients. We then use thewavelet basis to gauge the quality of the image, calculate the amount of blurriness,and finally correct the blur by removing it.Removing blur is the most important step towards making a high-quality three-dimensional image. This blur contains two main contributions: motion blur and op-160Figure 7.9: Processed iSCAT images of detached Hela cells. Left: Three-dimensional structure of fivedetached HeLa cells, with their nuclei in red. Right: An xz section of the left-hand image, with nucleiin red, and the rest sections of cell in blue.tical blur. This work discusses an efficient method to remove optical blur. Blurringmostly affects edges. Thus, considering the types of edges in an image provides ameasure of the extent of blurriness. We use this to assess the quality of an image,and find and remove the out-of-focus features. One of the main challenges thatedge detection techniques face is the identification of edges in noisy image areas.Edges are defined as any change in intensity between adjacent pixels; in general,an edge can be any small change in intensity in a pixel neighborhood, which canbe represented as a peak in the gradient domain. High levels of noise can disruptthe process of edge detection by confounding edge gradients with random intensitygradients. Wavelet transform is well-known for its adaptability to different noiselevels. Therefore, we use wavelet transform to investigate the type of edges in animage.Comparing the left- and right-hand images in Figure 7.7 illuminates the ef-161fect of wavelet deconvolution in removing noise and out-of-focus features. Theright-hand image is the sum of the fifteen treated focal planes, and shows that theapplied deconvolution technique has improved the contrast and the quality of im-age significantly. Figures 7.8 and 7.9 show the deconvoluted images of HeLa cellsat two different scales and with two different sample preparations. In Figure 7.8,cells are affixed to the glass coverslip, while in Figure 7.9, they are loose, resultingin two different structures. Note the higher level of detail in the image of affixedHeLa cells (Fig. 7.8). Figure 7.9 b represents a z cross section of unattached HeLacells, demonstrating the equal image quality and contrast for all focal planes acrossdifferent depths [23, 24, 190, 358].7.4.2 Gold Nanoparticle TrackingIt has become important in cellular biology to investigate the dynamics and mecha-nism of nanoparticle endocytosis. The following section describes the capability ofinterferometric scattering microscopy (iSCAT) in the context of particle tracking.Figure 7.5 displays time dependent images of gold nanoparticle internalizing in theHeLa cell.We investigated the interaction of 30 nm gold nanoparticles with HeLa cells,as a model cancer cell. File 2 in the supplementary information shows a videoof 30 nm gold nanoparticles interacting with different parts of a HeLa cell. Thevideo shows a gold nanoparticle attached to cell membrane, and then moving inthree dimensions within the cell. Our goal in this study was to investigate the iS-CAT’s capacity to capture time-dependent events, so the whole process of enteringnanoparticle into cells was not tracked rigorously.Figure 7.5 shows the iSCAT images of HeLa cells cultured with 30 nm goldnanoparticles. Bright circular spots in the image represent the gold nanoparticles.Particles that are in focus resemble white spots, while out-of-focus particles closeto focal plane appear as rings with a small spot in the center. These rings areAiry disks, caused by the point spread function. The center of mass of the goldnanoparticles is easily determined in most cases. To measure the center of massof the nanoparticles with higher accuracy, we fit the nanoparticles signal to a two-dimensional Gaussian function. External fluctuations, shot noise, or background162Figure 7.10: Three-dimensional gold nanoparticle trajectory while interacting with the exterior of a HeLacell membrane (100 points, 33 ms per point), reconstructed from the time-resolved iSCAT imageswith a temporal resolution of 33 ms.Figure 7.11: Three-dimensional particle trajectory while (left) passing through the membrane and (right)moving within the HeLa cell (1,200 points, 33 ms per point), both reconstructed from the time-resolved iSCAT images with a temporal resolution of 33 msec. The scalebar is 200nm.noise - such as out-of-focus signal contributions - cause a less-defined center ofmass. Thus, we need to remove all the confounding parameters before being ableto reliably track single particles. As we discussed in the previous section, wavelettransform subtracts the noise and out-of-focus features, while background subtrac-tion eliminates background from raw iSCAT images.We focused on one particular particle of interest to study time-resolved particle163Figure 7.12: HeLa cell after culturing with 30 nm gold nanoparticles for 4 hours. This image is a sum of12 different z sections.motion. Three time-resolved interactions of nanoparticles with the HeLa cell are ofinterest: interactions before, during, and after cell internalization. Figures 7.11 and7.10 illustrate the three-dimensional trajectories of nanoparticles, providing infor-mation of the type of the motion occurring. Each shows the motion of a separatenanoparticle at different stages of endocytosis.7.4.3 Dynamics After Adding Gold Nanoparticles to Cell SamplesThree-dimensional information about nanoparticle concentration and the localiza-tion of single or aggregated nanoparticles is important for the development of rel-evant nanoparticle applications, such as drug delivery or cell labeling. For thepresent study, we incubated HeLa cells with gold nanoparticles for four hours. Thecells were visualized by iSCAT microscopy at different depths within the sample.Figures 7.12 and 7.13 shows that gold nanoparticles are easily detected and identi-fied in a z-axis scan of a cell.Figure 7.12 shows HeLa cells after 4 hours of adding gold nanoparticles. Nanopar-ticles are visible as yellow-orange bright spots. To identify the exact position ofnanoparticles inside the cell, we took a z scan of iSCAT images over different cel-164Figure 7.13: Four different z sections of HeLa cells after culturing with 30 nm GNP for 4 hours. the depthis 8 µm with ∆z as 2 µmlular depths. Figure 7.13 shows iSCAT images of nanoparticle-infused HeLa cellsat different depths, showing that most of the nanoparticles are attached to the sur-face of the cell; however, some of nanoparticles are seen inside the cells, withoutbeing internalized in the nuclei.7.4.4 Raman SpectraIn Raman mapping, each pixel contains a large amount of spectral data, typi-cally 1,340 points, depending on the size of the spectrometer’s detector. Differentmethods are available to convert spectral intensities to a single pixel intensity; inother words, to reduce the dimensionality of the dataset from M×N× 1,340 toM×N×1. One efficient technique is Principal Component Analysis (PCA).The left-hand image in Figure 7.14 shows the first principal component (PC-1)of the Raman dataset of a sample containing both HeLa and fibroblast cells next165Figure 7.14: Interactions of gold nanoparticles with HeLa (lower left) and fibroblast (right) cells. Left:Map of first principal component of Raman data; orange areas are the most highly correlated withgold nanoparticles. Right: iSCAT image showing cells, before gold nanoparticles were added.to each other, after adding gold nanoparticles to the sample. The right-hand imageis the corresponding iSCAT image, before adding gold nanoparticles. The figureillustrates that Raman spectroscopy differentiates between not only different partsof individual cell, but also between healthy and cancerous cells. The underlyingreason Raman has this ability is that the ratio of nucleus to cytoplasm is larger incancerous cells compared to healthy cells. Moreover, cancerous cells have highermetabolic activity, causing differences in lipid and protein levels between healthyand cancerous cells.The orange areas of the Raman PC-1 map in Figure. 7.14 clearly reveals thatcancerous HeLa cells uptake gold nanoparticles more readily than healthy fibrob-last cells, due to higher permeability and retention effect of effect of cancerouscells.[135]7.4.5 Comparison Between iSCAT-Raman and Other TechniquesA variety of optical imaging techniques are available for the study of biologicalsystems. A brief review of common techniques alongside their advantages andlimitations is presented to help guide our discussion about the unique capabilitiesof iSCAT-Raman.166As we mentioned in the introduction, bright-field microscopy offers simplestapproach, and is used widely in biology. The low image contrast of biological sam-ples and low resolution of bright-field microscopy limit its applications in this areaof science, where researchers are increasingly interested in probing sub-cellularstructures. Phase-contrast and differential interference contrast (DIC) microscopyprovide fine details of samples; however, these two techniques are not suitable forthicker samples. Transmission electron microscopy (TEM)-tomography is a high-resolution technique capable of achieving 0.2 nm resolution, and able to create athree-dimensional image reconstruction of a cell. The need for thin sections ofsample and sample fixation restrict the applications of this technique.[5, 25, 126]Scanning electron microscopy (SEM), capable of achieving resolutions below 1nm, reveals details about the surface of samples with a good depth of field. SEMrequires that samples be coated with a conductive material. This process can distortcellular features by dehydrating the sample, and along with TEM, is unsuitable forin vivo imaging.[70, 192, 305].As we discussed, the aforementioned techniques require complex sample prepa-ration. Such preparatory procedures can affect the structure of biological samples,and create artifacts. It is not possible to study any reaction dynamics in a cellusing these techniques. The other shortcoming of most of the aforementioned mi-croscopy techniques is their lack of ability to provide three-dimensional image re-construction without very thin samples, or the preparation of different thin slices ofthe sample. To deal with these limitations, recent fluorescence-based approacheshave been widely used in the last decade.Fluorescence microscopy represents the most well-known category of high-resolution optical imaging techniques. A variety of different modified fluores-cence techniques are available, which produce high-resolution wide field three-dimensional images. However, all fluorescence methodologies suffer from theinherent need to label samples. The use of fluorophore-conjugated antibodies orfluorescent chemical dyes generally require preparative treatments that kill cells,making fluorescence techniques unsuitable for in vivo experiments. While studiesusing fluorescent proteins are well-suited to live cell studies, introduction of thefluorescent protein may disrupt native protein function, and requires the use of atransformable organism. One must also be concerned with the efficiency of the167labeling technique, as an insufficient density of fluorophores in the sample can in-troduce artifacts. Moreover, all fluorescence-based techniques are prone to photo-bleaching, a process in which prolonged excitation of a fluorescent marker causesits degradation.Setting these limitations aside, fluorescent markers are single quantum emit-ters, and can only emit a certain number of photons per unit time. In order foremission to occur, the fluorophore must populate an excited electronic state. Onthe other hand, photochemical effects limit the yield of emission photons. Theselimitations increase time necessary to observe a single emitter, to the order of aminute. For practical fluorescence microscopy, the achievable speed and localiza-tion of a single molecule roughly follows this equation:σ(time,space) = 1nm Hz−1/2 (7.5)If high localization accuracy is desirable, imaging at high speeds is ineffec-tive. Therefore, this trade-off limits the utility of fluorescence-based techniquesfor studying the dynamics of many biosystems, which occur faster than a certainrate.Extensive research has been carried out to develop an optical technique thatovercomes the limitations inherent to fluorescence-based techniques. Scatteringmicroscopy is a useful approach, as the incident photon flux is the only limitingfactor for detection and localization. Two scattering microscopies are of interest:pure scattering-based, and interferometric scattering-based techniques. In contrastto fluorescence microscopy, the photon flux in scattering techniques is only re-stricted by the amount of incident power on the sample, mitigated by scatteringcross section of the sample.In scattering-based techniques, we can tune the power of the incident light ac-cording to desired scattered photon flux. This makes it theoretically possible toachieve unlimited temporal and spatial resolution. However, in a practical experi-ment, background scattering limits the signal to noise ratio and consequently limitsdetection and localization.In general, the number of photons detected by the detector in scattering tech-nique obeys the following equation:168N = Ni[ζ 2+ |s|2+2ζ |s|cos∆φ] (7.6)where s is the scattering amplitude, Ni is the number of photons in incident light,∆φ refers to the phase difference between reference and scattered electric field, andζ refers to background (reference term), which is suppressed for dark field images.For objects with smaller diameter than the wavelength of the incident light, thescattering amplitude is a function of the size of the object as follows:s = |s|exp iφs (7.7)where s is described as:s = γεm(λ )piD32λ 2[εp(λ )− εm(λ )εp(λ )+2εm(λ )](7.8)in which γ is a proportionality constant, εm is the complex-valued dielectric con-stant of the medium, εp is the complex-valued dielectric constant of the particle,and λ is the incident wavelength.[229]The detected signal for pure scattering depends mainly on the pure scatteringterm |s|, while for interferometric scattering, the reference term ζ 2 dominates thesignal.In instrumental setups based on pure scattering techniques, such as dark-fieldmicroscopy, the detector is mounted perpendicular to the light source. In suchsetups, the only light detected by the detector is the scattered light; the setup rejectsincident illumination light, due to the detector’s perpendicular orientation.Instrumental background suppression in dark-field microscopy makes the spec-imen scattering term |s| dominate the detected signal, while background term ζ 2dominates in interferometric scattering. The remaining terms are negligible fordark-field; however, the third (interference) term in Eq. 7.6 becomes importantfor interferometric scattering, when the scattering fields of the sample and back-ground interfere. This happens when the optical path length difference betweenthe reference background and sample is smaller than the coherence length of theillumination light.Therefore, we can simplify Eq. 7.6 for the two scattering approaches, as fol-169lows:Dark-field (pure scattering):N ≈ Ni|s|2 (7.9)iSCAT (interferometric scattering):N ≈ Ni[ζ 2+2ζ |s|cos∆φ] (7.10)In interferometric scattering, the background term ζ 2 contains no informationabout the sample, so the interference term alone is of interest. Therefore, the sig-nal produced by the sample is observed as a small variation on top of the largebackground. The ratio of the signal to the background, Contrast, can be describedas:C =IsignalIbg(7.11)For pure scattering:C ≈ |s|2ζ(7.12)For interferometric scattering:C ≈ 2 |s|cos∆φζ(7.13)Note that for interferometric scattering, the signal contrast C linearly scaleswith the scattering amplitude |s|, and consequently to the third order of particlediameter (C ∝ D3; cf. Eq. 7.8). However, the dependence of contrast on particlesize increases to the sixth order for pure scattering (C ∝ D6).This signifies a big difference between interferometric scattering and eitherdark-field or fluorescence. The biggest limitation of pure scattering-based tech-niques is the high dependency on the size of the object (C ∝ D6). The applicationsof pure scattering-based techniques are restricted to samples larger than 200 nm,due to this dependency. Interferometric scattering improves this limitation anddecreases the order of the dependency. With interferometric techniques, the back-170ground is intense, compared to sample features. Fortunately, the background isalmost constant and can be removed by differential imaging.In this study, we used wide field iSCAT microscopy and removed the out-of-focus features using wavelet deconvolution techniques. The resulting deconvolutediSCAT image is comparable in quality to a confocal iSCAT image. However, thespeed of image acquisition is much faster than with confocal iSCAT. Another bene-fit of deconvoluted iSCAT is that it can be achieved at a lower incident illuminationpower, makes the technique better-suited to image light-sensitive specimens.7.5 ConclusionWe have presented and evaluated efficiency of the combination of interferometricscattering microscopy (iSCAT) and confocal Raman microscopy, by studying theinteraction of gold nanoparticles and human cells. Two types of cells, fibroblastsas model healthy cells and HeLa cells as model cancer cells, were investigated.Based on the results, iSCAT-Raman is able to provide real-time morphological andchemical information in three dimensions without the need for sample labels. Theadvantages of proposed technique over current spectroscopy techniques was alsodiscussed. Although an increasing number of techniques have been developed forbiological applications, none of them has proven capable of taking online, three-dimensional, label-free images with the ability to provide chemical information forevery area of a sample of interest. As we reviewed, fluorescence techniques are ef-ficient in taking fast, three-dimensional images with high resolution. However, theinherent need for labels limits the applications of fluorescence microscopy. Othermentioned techniques suffer from low resolution, poor contrast, long acquisitiontime, or restrictions in the thickness of a sample.171Chapter 8ConclusionA wealth of information creates a poverty of attention.Herbert SimonThis thesis has presented some novel chemometric methods to improve quan-titative and qualitative analysis by means of spectro/microscopic measurements.In the first part, chapters 3-6, it was demonstrated that Raman and NIR spectra ofunrefined pulp sample could accurately predict standard properties of paper madefrom that pulp at any arbitrary refining energy when appropriate data mining tech-niques were applied. In the second part, chapter 7, the efficiency of data miningtechniques were tested against images of biological cells. To do so, a new com-bination of iSCAT-Raman with data mining was introduced which was capableof taking online, three-dimensional, label-free images with the ability to providechemical information for every sample area of interest. This study was able toprove the success of this methodology in studying the dynamic of gold nanoparti-cle uptake in cancer and healthy cell models. The summary and conclusions of thefirst part are highlighted below• The first section showed that NIR was able to determine the physical prop-erties of paper sheet efficiently. We investigated the effectiveness of severalpreprocessing methods to determine which one could improve the multivari-ate prediction of paper physical and morphological properties from the NIRspectra of pulp fibres. It was found that a combination of the OSC and DWTresults in the smallest error. NIR spectra corrected by OSC-DWT revealed172interpretable features with respect to selected properties, enabling us to con-nect certain properties with assigned features. These results underlined theeffectiveness of OSC-DWT in improving the utility of NIR spectroscopy forpredicting the end-point properties of pulp and paper.• It was demonstrated that Raman spectra of unrefined pulp samples could ac-curately predict standard properties of paper made from that pulp at any arbi-trary refining energy, when modeled to a reduced space selected by a sophis-ticated chemometric feature selection technique referred to as a TemplateOriented Genetic Algorithm (TOGA). The results showed that TOGA couldefficiently identify and quantify the Raman spectra of systems of analyteswith overlapping spectral features and uncorrelated variance. TOGA re-moved irrelevant information in the spectrum and reduced the dimensional-ity of the calibration space. By isolating robustly correlated features, TOGAalso provided a spectrum of the covariance for chemical interpretation interms of functional group frequencies. We showed how the application ofTOGA to the Raman spectrum of unrefined pulps enabled the robust predic-tion of pulp network physical and mechanical properties for any arbitrarilyspecified level of pulp refining. TOGA showed that the physical character-istics of single fibres had the greatest influence on the mechanical propertiesof the unbeaten pulp or with low beating energy. On the other hand, TOGAsuccessfully identifies the hydrogen bond region as an essential part of thespectrum for classifying the mechanical properties of beaten pulp, since inhigher beating energy fibre interactions became a dominant factor.• Finally, we compared the effectiveness of NIR and Raman spectroscopytechniques in predicting pulp properties by applying the root mean squareerror of prediction to evaluate PLS models. It was found that NIR spec-troscopy performed better in predicting most of the mechanical properties ofpaper sheets. However, NIR spectra present only overlapping spectra withlimited information available, so does not offer an effective technique forspectroscopic troubleshooting. Raman spectroscopy on the other hand ex-tracts chemically specific signatures from aqueous samples with little or nosample preparation, compared with NIR techniques showed an overwhelm-173ing broad peak for water. We also found that Raman spectroscopy predictsproperties with higher beating and lower freeness more effectively than NIR.Results obtained by fusing these two techniques showed a slight improve-ment in the prediction model over the individual methods.The application of vibrational spectroscopy together with the development ofefficient data mining algorithms was not limited in pulp and paper industry. Byremoving uncorrelated features and highlighting the correlated ones in spectra uti-lizing data mining algorithms, vibrational spectroscopy can be justified as a strongonline tool in many industries. This claim was examined in the second part, chapter7. The summary and conclusions are as follows• We presented and evaluated efficiency of the combination of interferometricscattering microscopy (iSCAT) with confocal Raman microscopy in instru-mental part and efficiency of DWT in deconvolution part, by studying theinteraction of gold nanoparticles and human cells. Two types of cells, fi-broblasts as model healthy cells and HeLa cells as model cancer cells, wereinvestigated. Based on the results, iSCAT-Raman combined with DWT wasable to provide high resolution, real-time morphological and chemical infor-mation in three dimensions without the need for sample labels. It was shownthe fluorescence techniques were efficient in taking fast, three-dimensionalimages with high resolution. However, the inherent need for labels limitedthe applications of fluorescence microscopy. Other possible techniques suf-fer from low resolution, contrast, long acquisition time, or restrictions in thethickness of a sample.• The discrete wavelet transform was successfully utilized in denoising, reso-lution enhancement, and three-dimensional image processing of iSCAT im-ages. To accurately track a gold nanoparticle, we applied an algorithm basedon the two dimensional, Gaussian-corrected point spread function of a goldnanoparticle to accurately localized its center of its mass. The gold nanopar-ticle tracking study illustrated that the internalization of gold nanoparticlescould take more than an hour, due to the slow endocytosis process of the cell.• We showed that Raman spectroscopy not only differentiates between dif-174ferent parts of an individual cell, but also between healthy and cancerouscells. The results of iSCAT-Raman inquiry showed that cancerous HeLacells could uptake gold nanoparticles more readily than healthy fibroblastcells, likely due to higher permeability and retention effect of the cancer-ous taxonomy. Most of the nanoparticles are attached to the surface of thecell. But some of nanoparticles were seen inside the cells, without beinginternalized in the nuclei.Utilization of iSCAT-Raman combined with 2D data mining algorithms can po-tentially be expanded to different research areas such as drug delivery, biosens-ing, imaging, catalysis, and more due to the unique capability of this approach inproviding real-time morphological and chemical information in three dimensionswithout the need for sample labels.175Bibliography[1] H. Abramczyk and B. Brozek-Pluska. Raman imaging in biochemical andbiomedical applications. diagnosis and treatment of breast cancer.Chemical Reviews, 113(8):5766–5781, Aug. 2013.[2] M. J. Adams. Chemometrics in analytical spectroscopy. The Royal Societyof Chemistry, 1995.[3] J. Adur, G. Barbosa, V. Pelegati, M. Baratti, C. Cesar, V. Casco, andH. Carvalho. Multimodal and non-linear optical microscopy applications inreproductive biology. Microscopy Research and Technique, 79(7):567–582,2016.[4] U. P. Agarwal. Characterization of Lignocellulose Materials, chapterRaman Spectroscopic Characterization of Wood and Pulp Fibers, pages17–35. Blackwell Publishing, Oxford, UK, 2008.[5] K. T. Al-Jamal, H. Nerl, K. H. Mller, H. Ali-Boucetta, S. Li, P. D. Haynes,J. R. Jinschek, M. Prato, A. Bianco, K. Kostarelos, and A. E. Porter.Cellular uptake mechanisms of functionalised multi-walled carbonnanotubes by 3D electron tomography imaging. Nanoscale,3(6):2627–2635, 2011.[6] E. Alarousu, L. Krehut, T. Prykari, and R. Myllyla. Study on the use ofoptical coherence tomography in measurements of paper properties. Meas.Sci. Technol., 16:1131–1137, 2005.[7] B. aldini, E. Grilli, and M. Guzzi. Simple derivative optical spectrometer.Applied Optics, 14:2687–2690, 1975.[8] S. S. Alex Small. Fluorophore localization algorithms for super-resolutionmicroscopy. Nature Methods, 11:267– 279, 2014.176[9] J. B. Allen and L. R. Rabiner. A unified approach to short-time fourieranalysis and synthesis. Proceedings of the IEEE, 65(11):1558–1564, Nov1977.[10] J. R. Allen, S. T. Ross, and M. W. Davidson. Single molecule localizationmicroscopy for superresolution. Journal of Optics, 15(9):094001, Sept.2013.[11] R. D. Allen, N. S. Allen, and J. L. Travis. Video-enhanced contrast,differential interference contrast (AVEC-DIC) microscopy: A new methodcapable of analyzing microtubule-related motility in the reticulopodialnetwork of allogromia laticollaris. Cell Motility, 1(3):291–302, 1981.[12] B. K. Alsberg, A. M. Woodward, and D. B. Kell. An introduction towavelet transforms for chemometricians: A time-frequency approach.Chemometrics and Intelligent Laboratory Systems, 37(2):215 – 239, 1997.[13] A. C. Andersson. Direct orthogonalization. Chemometrics and IntelligentLaboratory Systems, 47:51–63, 1999.[14] M. R. Arnison, K. G. Larkin, C. J. R. Sheppard, N. I. Smith, and C. J.Cogswell. Linear phase imaging using differential interference contrastmicroscopy. Journal of Microscopy, 214(1):7–12, Apr. 2004.[15] J. O. Arroyo, J. Andrecka, K. M. Spillane, N. Billington, Y. Takagi, J. R.Sellers, and P. Kukura. Label-free, all-optical detection, imaging, andtracking of a single protein. Nano Letters, 14:2065–2070, 2014.[16] S. A. Asher. UV resonance raman spectroscopy for analytical, physical,and biophysical chemistry. Analytical Chemistry, 65(4):201–210, Feb.1993.[17] S. A. Asher, R. W. Bormett, X. G. Chen, D. H. Lemmon, N. Cho,P. Peterson, M. Arrigoni, L. Spinelli, and J. Cannon. Uv resonance ramanspectroscopy using a new cw laser source: Convenience and experimentalsimplicity. Applied Spectroscopy, 47(5):628–633, 1993.[18] E. Atilgan and B. . Ovryn. Reflectivity and topography of cells grown onglass-coverslips measured with phase-shifted laser feedback interferencemicroscopy. Biomed. Opt. Express, 2:24172437, 2011.[19] B. Bacskai. Multiphoton imaging of structure and function in mousemodels ofalzheimer’s disease. In Biomedical Optics 2016, page BTh3D.1.Optical Society of America, 2016.177[20] E. Bailo and V. Deckert. Tip-enhanced raman scattering. Chemical SocietyReviews, 37(5):921–930, 2008.[21] R. M. Balabina and S. V. Smirnov. Variable selection in near-infraredspectroscopy: Benchmarking of feature selection methods on biodieseldata. Anal. Chim. Acta, 29:63–72, 2011.[22] R. M. Balabina and S. V. Smirnov. Variable selection in near-infraredspectroscopy: Benchmarking of feature selection methods on biodieseldata. Anal. Chim. Acta, 29:63–72, 2011.[23] A. Banerjee, A. Berezhkovskii, and R. Nossal. Kinetics of cellular uptakeof viruses and nanoparticles via clathrin-mediated endocytosis. Phys. Biol.,13(1):016005, Feb. 2016.[24] A. M. Bannunah, D. Vllasaliu, J. Lord, and S. Stolnik. Mechanisms ofnanoparticle internalization and transport across an intestinal epithelial cellmodel: Effect of size and surface charge. Mol. Pharmaceutics,11(12):4363–4373, Dec. 2014.[25] M. Ba´rcena and A. J. Koster. Electron tomography in life science.Seminars in Cell & Developmental Biology, 20(8):920–930, Oct. 2009.[26] V. J. Barclay, R. F. Bonner, and I. P. Hamilton. Application of wavelettransforms to experimental spectra: Smoothing, denoising, and data setcompression. Analytical chemistry, 69:78–90, 1997.[27] A. Barnes and W. Orville. Vibrational SpectroscopyModern Trends.Elsevier Sci, Amsterdam, Holland, 1977.[28] G. R. Barnes, R.B. and U. Liddel. Infrared Spectroscopy. IndustrialApplications and Biography,. Reinhold Publishing Co,, New York, 1994.[29] R. J. Barnes, M. S. Dhanoa, and J. S. Lister. Standard normal variatetransformation and de-trending of near-infrared diffuse reflectance spectra.Applied Spectroscopy, 43:772–777, 1989.[30] M. Baroni, S. Clementi, G. Cruciani, G. Costantino, D. Riganelli, andE. Oberrauch. Predictive ability of regression models. part ii: Selection ofthe best predictive pls model. Journal of chemometrics, 6:347–356, 2005.[31] A. Barth. Infrared spectroscopy of proteins. Biochimica et Biophysica Acta(BBA) - Bioenergetics, 1767(9):1073 – 1101, 2007.178[32] R. F. Begley, A. B. Harvey, and R. L. Byer. Coherent anti-stokes ramanspectroscopy. Applied Physics Letters, 25(7):387–390, 1974.[33] A. J. Berger, Y. Wang, and M. S. Feld. Rapid, noninvasive concentrationmeasurements of aqueous biological analytes by near-infrared ramanspectroscopy. APPLIED OPTICS, 35:209–212, 1996.[34] M. Birkett and M. Gambino. Estimation of pulp kappa number withnear-infrared spectroscopy. TAPPI Journal, 72:193–197, 1989.[35] L. Bokobza. Near infrared spectroscopy. Journal of Near InfraredSpectroscopy, 6(1):3–17, 1998.[36] J. Borch. Handbook of Physical Testing of Paper, volume 1. CRC Press,2001.[37] E. Borrs, J. Ferr, R. Boqu, M. Mestres, L. Acea, and O. Busto. Data fusionmethodologies for food and beverage authentication and quality assessmenta review. Analytica Chimica Acta, 891:1 – 14, 2015.[38] D. Boyer. Photothermal imaging of nanometer-sized metal particles amongscatterers. Science, 297:11601163, 2002.[39] N. Bozinovic, C. Ventalon, T. Ford, and J. Mertz. Fluorescenceendomicroscopy with structured illumination. Opt. Express,16(11):8016–8025, May 2008.[40] R. Bro, A. Rinnana, and N. K. M. Faber. Standard error of prediction formultilinear pls 2. practical implementation in fluorescence spectroscopy.Chemometrics and Intelligent Laboratory Systems, 75:69–76, 2005.[41] F. K. Brown. Chapter 35 - chemoinformatics: What is it and how does itimpact drug discovery. volume 33 of Annual Reports in MedicinalChemistry, pages 375 – 384. Academic Press, 1998.[42] M. Brunner, R. Eugster, E. Trenka, and L. Bergamin-Strotz. Ft-nirspectroscopy and wood identification. Holzforschung - InternationalJournal of the Biology, Chemistry, Physics and Technology of Wood,50:130–134, 2009.[43] W. Cai, Y. Li, and X. Shao. A variable selection method based onuninformative variable elimination for multivariate calibration ofnear-infrared spectra. Chemometrics and Intelligent Laboratory Systems,90:188–194, 2008.179[44] F. Carrillo, X. Colom, J. Sun˜ol, and J. Saurina. Structural FTIR analysisand thermal characterisation of lyocell and viscose-type fibres. EuropeanPolymer Journal, 40(9):2229–2234, Sept. 2004.[45] P. Caspers, G. Lucassen, and G. Puppels. Combined in vivo confocalraman spectroscopy and confocal microscopy of human skin. BiophysicalJournal, 85(1):572–580, July 2003.[46] V. Centner and D. Massart. Elimination of uninformative variables formultivariate calibration. Analytical Chemistry, 68:3851–3858, 1996.[47] G. P. Chalmers JM. Handbook of Vibrational Spectroscopy,, volume 1.Chalmers JM, and Griffiths PR., UK, 2002.[48] F.-T. Chau, Y.-Z. Liang, J. Gaoo, and X.-G. Shao. Chemometrics: frombasic to wavelet transform, chapter Chapter 2, One dimensional signalprocessing techniques in chemiatryG, pages 23–69. Wiley, 2004.[49] D. Chen, Z. Chen, and E. Grant. Adaptive wavelet transform suppressesbackground and noise for quantitative analysis by raman spectrometry.Anal Bioanal Chem, 400:625–634, 2011.[50] D. Chen, B. Hu, X. Shao, and Q. Su. Variable selection by modified ipw(iterative predictor weighting)-pls (partial least squares) in continuouswavelet regression models. Analyst, 129:664–669, 2004.[51] D. Chen, T. Trung, H.-F. Jang, D. W. Francis, and E. R. Grant. Theprospect of raman spectroscopy as a gauge for process analysis / processcontrol in pulp and paper production. In Proc.Technical Association of thePulp and Paper Industry (TAPPI) Pulping, Engineering, Environmental,Recycling and Sustainability (PEERS) Conference, 2010.[52] D. Chen, T. Trung, H.-F. Jang, D. W. Francis, and E. R. Grant.High-throughput prediction of physical and mechanical properties of paperfrom raman chemometric analysis of pulp fibres. Canadian Journal ofForest Research, 41:2100–2113, 2011.[53] Z. Chen, T. Q. Hu, H. F. Jang, and E. Grant. Multivariate analysis ofhemicelluloses in bleached kraft pulp using infrared spectroscopy. AppliedSpectroscopy, Oct. 2016.[54] J.-X. Cheng, , and X. S. Xie. Coherent anti-stokes raman scatteringmicroscopy: instrumentation, theory, and applications. The Journal ofPhysical Chemistry B, 108:827–840, 2004.180[55] J. Choi, J. Choo, H. Chung, D.-G. Gweon, J. Park, H. J. Kim, S. Park, andC.-H. Oh. Direct observation of spectral differences between normal andbasal cell carcinoma (BCC) tissues using confocal raman microscopy.Biopolymers, 77(5):264–272, 2005.[56] I.-G. Chong and C.-H. Jun. Performance of some variable selectionmethods when multicollinearity is present. Chemometrics and IntelligentLaboratory Systems, 78:103–112, 2005.[57] A. Christy, N. Tavassoli, A. Bain, L. Melo, and E. R. Grant. Wide-fieldconfocal interferometric backscattering (iSCAT)-raman microscopy. InOptics in the Life Sciences, page NM4C.4. The Optical Society, 2015.[58] L. Chrit, C. Hadjur, S. Morel, G. Sockalingum, G. Lebourdon, F. Leroy,and M. Manfait. In vivo chemical investigation of human skin using aconfocal raman fiber optic microprobe. Journal of Biomedical Optics,10(4):044007, 2005.[59] R. J. H. Clark and T. J. Dines. Resonance raman spectroscopy, and itsapplication to inorganic chemistry. new analytical methods (27).Angewandte Chemie International Edition in English, 25(2):131–158,1986.[60] M. Clementi, S. Clementi, M. Fornaciari, F. Orlandi, and B. Romano. Thegolpe procedure for predicting olive crop production from climaticparameters. Journal of Chemometrics, 15:397–404, 2001.[61] C. J. Cogswell and C. J. R. Sheppard. Confocal differential interferencecontrast (DIC) microscopy: including a theoretical analysis of conventionaland confocal DIC imaging. Journal of Microscopy, 165(1):81–101, Jan.1992.[62] M. A. Coimbra, A. Barros, M. Barros, D. N. Rutledge, and I. Delgadillo.Multivariate analysis of uronic acid and neutral sugars in whole pecticsamples by ft-ir spectroscopy. Carbohydrate Polymers, 37(3):241 – 248,1998.[63] X. Colom and F. Carrillo. Crystallinity changes in lyocell and viscose-typefibres by caustic treatment. European Polymer Journal, 38(11):2225–2230,Nov. 2002.[64] X. S. X. Conor L. Evans. Coherent anti-stokes raman scatteringmicroscopy: Chemical imaging for biology and medicine. Annual Reviewof Analytical Chemistry, 1(1):883–909, July 2008.181[65] P. Cooper, D. Jeremic, S. Radivojevic, Y. Ung, and B. Leblon. Potential ofnear-infrared spectroscopy to characterize wood products 1 1 This article isa contribution to the series the role of sensors in the new forest productsindustry and bioeconomy. Canadian Journal of Forest Research,41(11):2150–2157, Nov. 2011.[66] T. M. Cotton, J.-H. Kim, and G. D. Chumanov. Application ofsurface-enhanced raman spectroscopy to biological systems. Journal ofRaman Spectroscopy, 22(12):729–742, Dec. 1991.[67] A. Curtis. The mechanism of adhesion of cells to glass a study byinterference reflection microscopy. J. Cell. Biol., 20:199215, 1964.[68] X.-d. Dai, B. Joseph, and R. L. Motard. Introduction to Wavelet Transformand Time-Frequency Analysis, pages 1–32. Springer US, Boston, MA,1994.[69] L. S. Davis. A survey of edge detection techniques. Computer Graphicsand Image Processing, 4(3):248 – 270, 1975.[70] W. Denk and H. Horstmann. Serial block-face scanning electronmicroscopy to reconstruct three-dimensional tissue nanostructure. PLoSBiology, 2(11):329, Oct. 2004.[71] M. Diem. Modern Vibrational Spectroscopy and Micro-Spectroscopy:Theory, Instrumentation and Biomedical Applications,. John Wiley &Sons, Ltd, first edition, 2015.[72] E. Dinc¸ and D. Baleanu. A review on the wavelet transform applications inanalytical chemistry, pages 265–284. Springer Netherlands, Dordrecht,2007.[73] O. Dopfer. Optical spectroscopy in chemistry and life sciences. anintroduction. by werner schmidt. ChemPhysChem, 7(7):1598–1598, 2006.[74] G. Downes, R. Meder, C. Hicks, and N. Ebdon. Developing and evaluatinga multisite and multispecies NIR calibration for the prediction of kraft pulpyield in eucalypts. Southern Forests: a Journal of Forest Science,71(2):155–164, June 2009.[75] E. Duesbury, J. Holliday, and P. Willett. Maximum commonsubstructure-based data fusion in similarity searching. Journal of ChemicalInformation and Modeling, 55(2):222–230, 2015. PMID: 25602464.182[76] P. Duhamel and M. Vetterli. Fast fourier transforms: A tutorial review anda state of the art. Signal Processing, 19(4):259 – 299, 1990.[77] E. Dusch, T. Dorval, N. Vincent, M. Wachsmuth, and A. Genovesio.Three-dimensional point spread function model for line-scanning confocalmicroscope with high-aperture objective. Journal of Microscopy,228(2):132–138, 2007.[78] M. ebek and P. Pta. The space variant psf for deconvolution of wide-fieldastronomical images, 2008.[79] C. Eggeling, A. Volkmer, and C. A. M. . Seidel. Molecular photobleachingkinetics of rhodamine 6g by one- and two-photon induced confocalfluorescence microscopy. ChemPhysChem, 6:791804, 2005.[80] F. Ehrentreich. Wavelet transform applications in analytical chemistry.Analytical and Bioanalytical Chemistry, 372:115–121, 2002.[81] E. S. Elzanfaly, S. A. Hassan, M. Y. Salem, and B. A. El-Zeany.Continuous wavelet transform, a powerful alternative to derivativespectrophotometry in analysis of binary and ternary mixtures: Acomparative study. Spectrochimica Acta Part A: Molecular andBiomolecular Spectroscopy, 151:945 – 955, 2015.[82] A. Enejder, T. Scecina, J. Oh, M. Hunter, W.-C. Shih, S. Sasic, G. L.Horowitz, and M. S. Feld. Raman spectroscopy for noninvasive glucosemeasurements. Journal of Biomedical Optics, 10:031114 (1–9), 2005.[83] H. H. Espy. The mechanism of wet-strength development in paper - areview. TAPPI Journal, pages 90–99, 1995.[84] N. J. Everall. Confocal Raman microscopy: Performance, pitfalls, bestpractice. Appl Spectrosc, 63(9):245–262, Sept. 2009.[85] H. L. F. Wei, J. Huang. Variable selection and estimation inhigh-dimensional varying-coefficient models. Statistica Sinica.,21:1515–1540, 2011.[86] J. Fahrenfort. Attenuated total reflection. Spectrochimica Acta, 17(7):698 –709, 1961.[87] O. Faix. Fourier Transform Infrared Spectroscopy, pages 83–109. SpringerBerlin Heidelberg, Berlin, Heidelberg, 1992.183[88] P. Fardim, M. Ferreira, and N. Duran. Multivariate calibration forquantitative analysis of eucalypt kraft pulp by nir spectrometry. Journal ofWood Chemistry and Technology, 22:67–81, 2002.[89] P. Fardim, M. M. C. Ferreira, and N. Dura´n. Determination of mechanicaland optical properties of eucalyptus kraft pulp by NIR spectrometry andmultivariate calibration. Journal of Wood Chemistry and Technology,25(4):267–279, Oct. 2005.[90] P. J. Ferreira, J. A. Gamelas, I. M. Moutinho, A. G. Ferreira, N. Gmez,C. Molleda, and M. M. Figueiredo. Application of FT-IR-ATRspectroscopy to evaluate the penetration of surface sizing agents into thepaper structure. Industrial & Engineering Chemistry Research,48(8):3867–3872, Apr. 2009.[91] S. Fischer, K. Schenzel, K. Fischer, and W. Diepenbrock. Applications ofFT raman spectroscopy and micro spectroscopy characterizing celluloseand cellulosic biomaterials. Macromolecular Symposia, 223(1):41–56,Mar. 2005.[92] M. Forina, C. Casolinoa, and E. M. Almansa. The refinement of pls modelsby iterative weighting of predictor variables and objects. Chemometricsand Intelligent Laboratory Systems, 68:29–40, 2003.[93] M. Forina, S. Lanteri, M. C. C. Oliveros, and C. P. Millan. Selection ofuseful predictors in multivariate calibration. Analytical and BioanalyticalChemistry, 380:397–418, 2004.[94] A. G. Frenich, D. Jouan-Rimbaud, D. L. Massart, S. Kuttatharmmakul,M. M. Galera, and J. L. M. Vidal. Wavelength selection method formulticomponent spectrophotometric determinations using partial leastsquares. Analyst, 120:2787– 2792, 1995.[95] C. W. Freudiger, W. Min, B. G. Saar, S. Lu, G. R. Holtom, C. He, J. C.Tsai, J. X. Kang, and X. S. Xie. Label-free biomedical imaging with highsensitivity by stimulated raman scattering microscopy. Science,322(5909):1857–1861, 2008.[96] C.-C. Fu, H.-Y. Lee, K. Chen, T.-S. Lim, H.-Y. Wu, P.-K. Lin, P.-K. Wei,P.-H. Tsao, H.-C. Chang, and W. Fann. Characterization and application ofsingle fluorescent nanodiamonds as cellular biomarkers. Proc. Natl. Acad.Sci., 104:727732, 2007.184[97] P. K. Gabrielle de Wit, John S. H. Danial and M. I. Wallace. Dynamiclabel-free imaging of lipid nanodomains. PNAS, 112:12299–12303, 2015.[98] P. Garside and P. Wyeth. Identification of cellulosic fibres by ftirspectroscopy - thread and single fibre analysis by attenuated totalreflectance. Studies in Conservation, 48(4):269–275, 2003.[99] P. Geladi and R. B. Kowalski. Parpart least squares regression: a tutorial.Analytical Chimica Acta, 185:1–17, 1986.[100] P. Geladi, D. MacDougall, and H. Martens. Linearization andscatter-correction for near-infrared reflectance spectra of meat. AppliedSpectroscopy, 39:491–500, 1985.[101] S. Gharehkhani, E. Sadeghinezhad, S. N. Kazi, H. Yarmand, A. Badarudin,M. R. Safaei, and M. N. M. Zubir. Basic effects of pulp refining on fiberpropertiesa review. Carbohydrate Polymers, 115:785 – 803, 2015.[102] A. T. Giese and C. S. French. The analysis of overlapping spectralabsorption bands by derivative spectrophotometry. Applied Spectroscopy,9:78–96, 1955.[103] W. Gindl and A. Teischinger. The potential of vis- and nir-spectroscopy forthe nondestructive evaluation of grain-angle in wood. Wood and FiberScience, 34:651–656, 2002.[104] W. Gindl, A. Teischinger, M. Schwanninger, and B. Hinterstoisser. Therelationship between near infrared spectra of radial wood surfaces andwood mechanical properties. Journal of Near Infrared Spectroscopy,9:255–261, 2001.[105] J. V. Glenn, J. R. Beattie, , L. Barrett, N. Frizzell, S. R. Thorpe, M. E.Boulton, J. J. McGarvey, and A. W. Stitt. Confocal raman microscopy canquantify advanced glycation end product (age) modifications in bruchsmembrane leading to accurate, nondestructive prediction of ocular aging.The FASEB Journal, 21(13):3542–3552, Nov. 2007.[106] R. Gobel, R. Krska, R. Kellner, R. W. Seitz, and S. A. Tomellini.Investigation of different polymers as coating materials for ir/atrspectroscopic trace analysis of chlorinated hydrocarbons in water. Appl.Spectrosc., 48(6):678–683, Jun 1994.185[107] H. Goicoechea and A. Olivieri. A new family of genetic algorithms forwavelength interval selection in multivariate analytical spectroscopy.JOURNAL OF CHEMOMETRICS, 17:338–345, 2003.[108] K. C. Gordon and S. J. Fraser-Miller. Raman Spectroscopy, pages139–169. Springer New York, New York, NY, 2016.[109] D. Griffin and J. Lim. Signal estimation from modified short-time fouriertransform. IEEE Transactions on Acoustics, Speech, and SignalProcessing, 32(2):236–243, Apr 1984.[110] P. R. Griffiths, C. T. Foskett, and R. Curbelo. Rapid scan infrared fouriertransform spectroscopy. Applied Spectroscopy Reviews, 6(1):31–77, 1972.[111] R. Guchardi, P. A. da Costa Filho, R. Poppi, and C. Pasquini.Determination of ethanol and methyl tert-butyl ether (mtbe) in gasoline bynir aotf-based spectroscopy and multiple linear regression with variablesselected by genetic algorithm. Journal of Near Infrared Spectroscopy,6(1):333–339, 1998.[112] L. W. H. Antti, M. Sjstrm. Multivariate calibration models using NIRspectroscopy on pulp and paper industrial applications. Journal ofChemometrics, 10(5-6):591–603, Sept. 1996.[113] G. S. H. Yuan, T. Garver. Spectroscopic methods for monitoring pulpbleaching processes. Pulp Paper Can.,, 107:49–51, 2006.[114] D. M. Haaland and E. V. Thomas. Partial least-squares methods forspectral analyses. 1. relation to other quantitative calibration methods andthe extraction of qualitative information. Analytical chemistry,60:1193–1202, 1988.[115] R. N. Hager and J. R. C. a. Anderson. Theory of the derivativespectrometer. Journal of the optical society of America, 60:1444–1449,1970.[116] E. W. Hansen, J. Zelten, and B. A. Wiseman. Laser scanning fluorescencemicroscope. In J. R. Lakowicz, editor, Time-Resolved Laser Spectroscopyin Biochemistry, volume 0909, pages 304–311. SPIE-Intl Soc Optical Eng,June 1988.[117] R. K. Hanson, R. M. Spearrin, and C. S. Goldenstein. Rayleigh and RamanSpectra, pages 91–105. Springer International Publishing, Cham, 2016.186[118] T. Haraguchi. Live cell imaging: Approaches for studying proteindynamics in living cells. Cell Structure and Function, 27(5):333–334,2002.[119] A. Hartschuh, E. J. Sa´nchez, X. S. Xie, and L. Novotny. High-resolutionnear-field raman microscopy of single-walled carbon nanotubes. Phys. Rev.Lett., 90:095503, Mar 2003.[120] J. B. Hauksson, G. Bergqvist, U. Bergsten, M. Sjstrm, and U. Edlund.Prediction of basic wood properties for Norway spruce. interpretation ofnear infrared spectroscopy data using partial least squares regression. WoodScience and Technology, 35(6):475–485, Dec. 2001.[121] S. W. Hell and J. Wichmann. Breaking the diffraction resolution limit bystimulated emission: stimulated-emission-depletion fluorescencemicroscopy. Optics Letters, 19(11):780–782, June 1994.[122] T. Hellerer, C. Axang, C. Brackmann, P. Hillertz, M. Pilon, and A. Enejder.Monitoring of lipid storage in caenorhabditis elegans using coherentanti-stokes raman scattering (CARS) microscopy. Proceedings of theNational Academy of Sciences, 104(37):14658–14663, Sept. 2007.[123] R. W. Hellwarth. Theory of stimulated raman scattering. Phys. Rev.,130:1850–1852, Jun 1963.[124] H. Henriksen, T. Naes, R. Rodbotten, and A. Aastveit. Simultaneousmodelling of process variables and raw material properties as measured bynir. a case study from cellulose production. Chemometrics and IntelligentLaboratory Systems, 77:238–246, 2005.[125] G. Herzberg. Molecular Spectra and Molecular Structure, vol. 1, Spectraof Diatomic Molecules. Van Nostrand,, New York, 1950.[126] K. Hhn, M. Sailer, L. Wang, M. Lorenz, M. Schneider, and P. Walther.Preparation of cryofixed cells for improved 3D ultrastructure with scanningtransmission electron tomography. Histochemistry and Cell Biology,135(1):1–9, Nov. 2010.[127] A. R. Hind, S. K. Bhargava, and A. McKinnon. At the solid/liquidinterface: Ftir/atr-the tool of choice. Advances in Colloid and InterfaceScience, 93(13):91 – 114, 2001.187[128] L. Hirvonen, K. Wicker, O. Mandula, and R. Heintzmann. Structuredillumination microscopy of a living cell. European Biophysics Journal,2009.[129] L. M. Hirvonen, K. Wicker, O. Mandula, and R. Heintzmann. Structuredillumination microscopy of a living cell. European Biophysics Journal,38(6):807–812, June 2009.[130] A. J. Hobro, J. Kuligowski, M. Dll, and B. Lendl. Differentiation of walnutwood species and steam treatment using ATR-FTIR and partial leastsquares discriminant analysis (PLS-DA). Analytical and BioanalyticalChemistry, 398(6):2713–2722, Sept. 2010.[131] G. Horlick. Introduction to fourier transform spectroscopy. Appl.Spectrosc., 22(6):617–626, Nov 1968.[132] A. a. Hoskuldsson. Pls regression method. Journal of Chemometrics,2:211–228, 1988.[133] B. Huang, M. Bates, and X. Zhuang. Super-resolution fluorescencemicroscopy. Annual Review of Biochemistry, 78(1):993–1016, June 2009.[134] H. Huang, H. Yu, H. Xu, and Y. Ying. Near infrared spectroscopy foron/in-line monitoring of quality in foods and beverages: A review. Journalof Food Engineering, 87(3):303 – 313, 2008.[135] X. Huang and M. A. El-Sayed. Gold nanoparticles: Optical properties andimplementations in cancer diagnosis and photothermal therapy. Journal ofAdvanced Research, 1(1):13–28, Jan. 2010.[136] G. R. Hunt. Spectral signatures of particulate minerals in the visible andnear infrared. GEOPHYSICS, 42(3):501–513, 1977.[137] A. A. Ibrahem, M. A. Yousef, and S. A. El-Meadawy. Effect of beating onfibre crystallinity and physical properties of paper sheets. Med J IslamicWorld Acad Sci, 2(4):295–298, 1989. doi:.[138] K. Ishimaru, T. Hata, P. Bronsveld, D. Meier, and Y. Imamura.Spectroscopic analysis of carbonization behavior of wood, cellulose andlignin. Journal of Materials Science, 42(1):122–129, Nov. 2006.[139] S. Iwamoto, K. Abe, and H. Yano. The effect of hemicelluloses on woodpulp nanofibrillation and nanofiber network characteristics.Biomacromolecules, 9(3):1022–1026, 2008. PMID: 18247566.188[140] V. Jacobsen, P. Stoller, C. Brunner, V. Vogel, and V. Sandoghdar.Interferometric optical detection and tracking of very small goldnanoparticles at a water-glass interface. Opt Express, 14:405–414, 2006.[141] R. M. Jarvis and R. Goodacre. Discrimination of bacteria usingsurface-enhanced raman spectroscopy. Analytical Chemistry, 76(1):40–47,Jan. 2004.[142] E. JC. in Infra-Red Spectroscopy and Molecular Structure. An Outline ofthe Principles. Elsevier, Amsterdam, 1963.[143] A. Jensen and A. la Cour-Harbo. Ripples in Mathematics: The DiscreteWavelet Transform. Springer, 2001.[144] J. L. Jianqing Fan. A selective overview of variable selection in highdimensional feature space. Mathematical Reviews, 20:101–148, 2010.[145] J. S. JJ Davenport, J Hodgkinson and R. Tatam. Noise analysis for ccdbased ultra-violet and visible spectrophotometry. Appl. Opt.,54:8135–8144, 2015.[146] H. Jodl. Vibrational Spectra and Structure. A series of advances,volume 13. Elsevier, Amsterdam, Holland,, 1984.[147] A. Johansonn. Correlations between fibre properties and paper properties.PhD thesis, KTH Royal Institute of Technology, Stockholm, 2011.[148] D. Jouan-Rimbaud, B. Walczak, R. J. Poppi, O. E. de Noord, and D. L.Massart. Application of wavelet transform to extract the relevantcomponent from spectral data for multivariate calibration. Analyticalchemistry, 69:4317–4323, 1997.[149] O. Joutsimo, R. Wathe´n, and T. Tamminen. Effects of fiber deformationson pulp sheet properties and fiber strength. Paper and Timber, 87, 2005.[150] J. J. W. JR. Review of process and non-invasive near-infrared and infraredspectroscopy:. Applied Spectroscopy Reviews, 34(1-2):1–89, 1999.[151] P. A. W. Jr. and T. Hirschfeld. Internal reflection spectroscopy. AppliedSpectroscopy Reviews, 1(1):99–130, 1967.[152] R. P. Judy, J. J. Keating, E. M. DeJesus, J. X. Jiang, O. T. Okusanya, S. Nie,D. E. Holt, S. P. Arlauckas, P. S. Low, E. J. Delikatny, and S. Singhal.Quantification of tumor fluorescence during intraoperative optical cancerimaging. Scientific Reports, 5:16208, Nov. 2015.189[153] B. M. Julian Haas. Advances in mid-infrared spectroscopy for chemicalanalysis. Annual Review of Analytical Chemistry, 9(1):45–68, 2016.PMID: 27070183.[154] U. Kaiser, J. Biskupek, J. Meyer, J. Leschner, L. Lechner, H. Rose,M. Stger-Pollach, A. Khlobystov, P. Hartel, H. Mller, M. Haider,S. Eyhusen, and G. Benner. Transmission electron microscopy at 20 kv forimaging and spectroscopy. Ultramicroscopy, 111(8):1239 – 1246, 2011.[155] G. Kalinkova. Infrared spectroscopy in pharmacy. VibrationalSpectroscopy, 19(2):307 – 320, 1999.[156] T. Kang and H. Paulapuro. Effect of external fibrillation on paperstrength.Pulp Paper Canada, 107:51–54, 2006.[157] F. Kerouh and A. Serir. A perceptual blind blur image quality metric. In2014 IEEE International Conference on Acoustics, Speech and SignalProcessing (ICASSP). Institute of Electrical and Electronics Engineers(IEEE), May 2014.[158] K. Klein, A. M. Gigler, T. Aschenbrenner, R. Monetti, W. Bunk,F. Jamitzky, G. Morfill, R. W. Stark, and J. Schlegel. Label-free live-cellimaging with confocal raman microscopy. Biophysical Journal,102(2):360–368, Jan. 2012.[159] D. Klemm, B. Heublein, H.-P. Fink, and A. Bohn. Cellulose: Fascinatingbiopolymer and sustainable raw material. Angew. Chem. Int. Ed.,44:3358–93, 2005.[160] K. Kneipp, A. S. Haka, H. Kneipp, K. Badizadegan, N. Yoshizawa,C. Boone, K. E. Shafer-Peltier, J. T. Motz, R. R. Dasari, and M. S. Feld.Surface-enhanced raman spectroscopy in single living cells using goldnanoparticles. Applied Spectroscopy, 56(2):150–154, Feb. 2002.[161] J. Koenig. Application of fourier transform infrared spectroscopy tochemical systems. Applied Spectroscopy, 29(4):293–308, 1975.[162] T. Kondo and C. Sawatari. A Fourier transform infra-red spectroscopicanalysis of the character of hydrogen bonds in amorphous cellulose.Polymer, 37(3):393–399, Feb. 1996.[163] P. KOZUBEK, M.; MATULA. An efficient algorithm for measurement andcorrection of chromatic aberrations in fluorescence microscopy. Journal ofMicroscopy, 200:206–217, 2000.190[164] R. S. Krishnan and R. K. Shankar. Raman effect: History of the discovery.Journal of Raman Spectroscopy, 10(1):1–8, 1981.[165] H. Kubinyi. Evolutionary variable selection in regression and pls analyses.Journal of chemometrics, 10:119–133, 1996.[166] U. Kubitscheck, O. Kckmann, and R. Kues, T. & Peters. Imaging andtracking of single gfp molecules in solution. Biophys. J., 78:21702179,2000.[167] T. Kues and U. Peters, R. & Kubitscheck. Visualization and tracking ofsingle protein molecules in the cell nucleus. Biophys. J., 80:29542967,2001.[168] M. F. Langhorst, J. Schaffer, and B. Goetze. Structure brings clarity:Structured illumination microscopy in cell biology. Biotechnology Journal,4(6):858–865, June 2009.[169] P. Lasch. Spectral pre-processing for biomedical vibrational spectroscopyand microspectroscopic imaging. Chemometrics and Intelligent LaboratorySystems, 117:100 – 114, 2012. Special Issue Section: Selected Papers fromthe 1st African-European Conference on Chemometrics, Rabat, Morocco,September 2010 Special Issue Section: Preprocessing methods SpecialIssue Section: Spectroscopic imaging.[170] R. Leardi. Application of genetic algorithm-pls for feature selection inspectral data sets. Journal of Chemometrics, 14:643–655, 2000.[171] R. Leardi and A. L. Gonzalez. Genetic algorithms applied to featureselection in pls regression: how and when to use them. Chemometrics andIntelligent Laboratory Systems, 41:195–207, 1998.[172] T. Lestander and C. Rhen. Multivariate nir spectroscopy models formoisture and ash and calorific content in biofuels using bi-orthogonalpartial least square regression. Analyst, 130:1182–1189, 2005.[173] C. Li, G. Xiong, Q. Xin, J. Liu, P. L. Ying, Z. Feng, J. Li, w. b. Yang, Y. Z.Wang, G. r Wang, X. Liu, M. Lin, X. Wang, and E. Min. Uv resonanceraman spectroscopic identification of titanium atoms in the framework ofts-1 zeolite. Angewandte Chemie International Edition, 38(15):2220–2222,1999.[174] E. Li-Chan, A. Ismail, J. Sedman, and F. van de Voort. VibrationalSpectroscopy of Food and Food Products. John Wiley & Sons, Ltd, 2006.191[175] J. S. Lim. Two-dimensional Signal and Image Processing. Prentice-Hall,Inc., Upper Saddle River, NJ, USA, 1990.[176] Y.-H. Lin and C. Chang, W.-L. & Hsieh. Shot-noise limited localization ofsingle 20 nm gold particles with nanometer spatial precision withinmicroseconds. Opt Express, 22:9159, 2014.[177] F. Lindgren, P. Geladi, S. Rannar, and S. Wold. Interactive variableselection (ivs) for pls. part i. theory and algorithms. Journal ofChemometrics, 8:349– 363, 1994.[178] T. Lindgren, U. Edlund, and T. Iversen. A multivariate characterization ofcrystal transformations of cellulose. Cellulose, 2(4):273–288, Dec. 1995.[179] R. Liu, Z. Li, and J. Jia. Image partial blur detection and classification. In2008 IEEE Conference on Computer Vision and Pattern Recognition, pages1–8, June 2008.[180] W.-J. Z. Lu Xu. Comparison of different methods for variable selection.Analytica Chimica Acta, 446:475–481, 2001.[181] C. Lucasius, M. Beckers, and G. Kateman. Genetic algorithms inwavelength selection: a comparative study. Analytica Chimica Acta,286:135–153, 1994.[182] C. Lucasius and G. Kateman. Understanding and using genetic algorithmspart 1. concepts, properties and context. Chemometrics and IntelligentLaboratory Systems, 19:1–33, 1993.[183] C. C. M. Forina and C. P. Millan. Iterative predictor weighting (ipw) pls: atechnique for the elimination of useless predictors in regression problems.Journal of chemometrics, 13:165–184, 1999.[184] I. MacLeod, A. Scully, K. Ghiggino, P. Ritchie, O. Paravagna, andB. Leary. Photodegradation at the wood-clearcoat interface. Wood Scienceand Technology, 29(3), May 1995.[185] L. S. Magwaza, U. L. Opara, H. Nieuwoudt, P. J. R. Cronje, W. W. Saeys,and B. Nicolaı¨. Nir spectroscopy applications for internal and externalquality analysis of citrus fruit—a review. Food and Bioprocess Technology,5(2):425–444, 2012.[186] M. Maier. Applications of stimulated raman scattering. Applied physics,11(3):209–231, 1976.192[187] K. Majzner, A. Kaczor, N. Kachamakova-Trojanowska, A. Fedorowicz,S. Chlopicki, and M. Baranska. 3D confocal raman imaging of endothelialcells and vascular wall: perspectives in analytical spectroscopy ofbiomedical research. The Analyst, 138(2):603–610, 2013.[188] J. Malinen, M. Knskoski, R. Rikola, and C. G. Eddison. Led-based {NIR}spectrometer module for hand-held and process analyser applications.Sensors and Actuators B: Chemical, 51(13):220 – 226, 1998.[189] S. Mallat. A Wavelet Tour of Signal Processing, Third Edition: The SparseWay. Academic Press, 3rd edition, 2008.[190] S. K. Mann, E. Czuba, L. I. Selby, G. K. Such, and A. P. R. Johnston.Quantifying nanoparticle internalization using a high throughputinternalization assay. Pharm Res, 33(10):2421–2432, July 2016.[191] M. D. Marchi, V. Toffanin, M. Cassandro, and M. Penasa. Invited review:Mid-infrared spectroscopy as phenotyping tool for milk traits. Journal ofDairy Science, 97(3):1171 – 1186, 2014.[192] F. Marinello, P. Bariani, E. Savio, A. Horsewell, and L. D. Chiffre. Criticalfactors in SEM 3D stereo microscopy. Measurement Science andTechnology, 19(6):065705, May 2008.[193] A. Marklund, M. Paper, J. Hauksson, and M. Sjostrom. Prediction ofstrength parameters for softwood kraft pulps. multivariate data analysisbased on orthogonal signal correction and near infrared spectroscopy.Nordic Pulp & Paper Research Journal, 14:140–148, 1999.[194] H. Martens, J. Nielsen, and S. Engelsen. Light scattering and lightabsorbance separated by extended multiplicative signal correction.application oto near-infrared transmission analysis of powder mixture.Analytical Chemistry, 75:394–404, 2003.[195] H. Martens and E. Stark. Extended multiplicative signal correction andspectral interference subtraction: New preprocessing methods for nearinfrared spectroscopy. Joiurnal of Pharmaceutical and BiomedicalAnalysis, 9:625–635, 1991.[196] M. Marvin. Microscopy apparatus, December 1961.[197] K. F. Masamoto Arakawa, Yosuke Yamashita. Genetic algorithm-basedwavelength selection method for spectral calibration. JOURNAL OFCHEMOMETRICS, 25:10–19, 2010.193[198] T. Matsuzaki, G. Sazaki, M. Suganuma, T. Watanabe, T. Yamazaki,M. Tanaka, S. Nakabayashi, and H. Y. Yoshikawa. High contrastvisualization of cellhydrogel contact by advanced interferometric opticalmicroscopy. J. Phys. Chem. Lett., pages 253–257, 2013.[199] W. F. McClure. Near-infrared spectroscopy. the giant is running strong.Analytical Chemistry, 66:43A–53A, 1994.[200] T. Mehmood, K. H. Liland, L. Snipen, and S. Sb. A review of variableselection methods in partial least squares regression. Chemometrics andIntelligent Laboratory Systems, 118:62 – 69, 2012.[201] J. Miao, Z. Luo, Y. Wang, and G. Li. Comparison and data fusion of anelectronic nose and near-infrared reflectance spectroscopy for thediscrimination of ginsengs. Anal. Methods, 8:1265–1273, 2016.[202] J. Miller and J. Miller. Statistics for analytical chemistry, 2nd edition. JohnWiley and Sons,New York, NY, Jan 1988.[203] M. Mller and A. Zumbusch. Coherent anti-stokes raman scatteringmicroscopy. ChemPhysChem, 8(15):2156–2170, 2007.[204] D. P. Moerner, W. E. & Fromm. Methods of single-molecule fluorescencespectroscopy and microscopy. Rev. Sci. Instrum., 74:3597, 2003.[205] I. Mohammed-Ziegler, Z. Ho´rvlgyi, A. To´th, W. Forsling, andA. Holmgren. Wettability and spectroscopic characterization of silylatedwood samples. Polymers for Advanced Technologies, 17(11-12):932–939,2006.[206] B. Mohebby. Attenuated total reflection infrared spectroscopy of white-rotdecayed beech wood. International Biodeterioration & Biodegradation,55(4):247–251, June 2005.[207] B. Mohebby. Application of atr infrared spectroscopy in wood acetylation.Journal of Agricultural Science and Technology, 10:253–259, 2010.[208] N. Mojarad, V. Sandoghdar, and M. Krishnan. Measuringthree-dimensional interaction potentials using optical interference. Opt.Express, 21(8):9377–9389, Apr 2013.[209] C. Monzel, S. F. Fenz, and K. Merkel, R. & Sengupta. Probingbiomembrane dynamics by dual-wavelength reflection interference contrastmicroscopy. ChemPhysChem, 10:28282838, 2009.194[210] C. R. Mora and L. R. Schimleckab. On the selection of samples formultivariate regression analysis: application to near-infrared (nir)calibration models for the prediction of pulp yield in eucalyptus nitens.Can. J. For. Res., 38:2626–2634, 2008.[211] C. R. Mora and L. R. Schimleckab. On the selection of samples formultivariate regression analysis: application to near-infrared (NIR)calibration models for the prediction of pulp yield in eucalyptus nitens.Canadian Journal of Forest Research, 38(10):2626–2634, Oct. 2008.[212] M. D. Morris and D. J. Wallan. Resonance raman spectroscopy. AnalyticalChemistry, 51:182A–192A, 1979.[213] M. Moskovits. Surface-enhanced raman spectroscopy: a briefretrospective. Journal of Raman Spectroscopy, 36(6-7):485–496, 2005.[214] M. F. Mrozek and M. J. Weaver. Detection and identification of aqueoussaccharides by using surface-enhanced raman spectroscopy. Analyticalchemistry, 74:4069–4075, 2002.[215] B. Mulloy, G. W. Hart, and P. Stanley. Structural analysis of glycans. InA. Varki, R. D. Cummings, J. D. Esko, H. H. Freeze, P. Stanley, C. R.Bertozzi, G. W. Hart, and M. E. Etzler, editors, Essentials of Glycobiology,chapter 47. Cold Spring Harbor Laboratory Press, 2009.[216] S. H. Nawab and T. F. Quatieri. Advanced topics in signal processing.chapter Short-time Fourier Transform, pages 289–337. Prentice-Hall, Inc.,Upper Saddle River, NJ, USA, 1987.[217] M. L. Nelson and R. T. O’Connor. Relation of certain infrared bands tocellulose crystallinity and crystal lattice type. part II. a new infrared ratiofor estimation of crystallinity in celluloses i and II. Journal of AppliedPolymer Science, 8(3):1325–1341, May 1964.[218] M. L. Nelson and R. T. O’Connor. Relation of certain infrared bands tocellulose crystallinity and crystal latticed type. part i. spectra of latticetypes i, II, III and of amorphous cellulose. Journal of Applied PolymerScience, 8(3):1311–1324, May 1964.[219] U. Neugebauer, P. Rsch, M. Schmitt, J. Popp, C. Julien, A. Rasmussen,C. Budich, and V. Deckert. On the way to nanometer-sized information ofthe bacterial surface by tip-enhanced raman spectroscopy.ChemPhysChem, 7(7):1428–1430, July 2006.195[220] A. Niazi and R. Leardi. Genetic algorithms in chemometrics. JOURNALOF CHEMOMETRICS, 26:345–351, 2012.[221] S. Nie. Probing single molecules and single nanoparticles bysurface-enhanced raman scattering. Science, 275(5303):1102–1106, Feb.1997.[222] S. Nioka and B. Chance. Nir spectroscopic detection of breast cancer.Technology in Cancer Research & Treatment, 4(5):497–512, 2005.[223] Y. S. Nirit Dudovich, Dan Oron. Single-pulse coherently controllednonlinear raman spectroscopy and microscopy. Letters to Nature,418:512–514, 2002.[224] Y. Nishiyama, P. Langan, and H. Chanzy. Crystal structure andhydrogen-bonding system in cellulose i from synchrotron x-ray andneutron fiber diffraction. Journal of the American Chemical Society,124(31):9074–9082, 2002. PMID: 12149011.[225] L. Nrgaard, A. Saudland, J. Wagner, J. P. Nielsen, L. Munck, and S. B.Engelsen. Interval partial least-squares regression (ipls): A comparativechemometric study with an example from near-infrared spectroscopy.Applied Spectroscopy, 54(3):413–419, 2000.[226] S. R. Ober, Raimund J. and E. S. Ward. Localization accuracy insingle-molecule microscopy. Biophys J, 86:1185–1200, 2004.[227] A. Oddone, I. V. Vilanova, J. Tam, and M. Lakadamyali. Super-resolutionimaging with stochastic single-molecule localization: Concepts, technicaldevelopments, and biological applications. Microscopy Research andTechnique, 77(7):502–509, Feb. 2014.[228] M. Oheim, D. J. Michael, M. Geisbauer, D. Madsen, and R. H. Chow.Principles of two-photon excitation fluorescence microscopy and othernonlinear imaging approaches. Advanced Drug Delivery Reviews,58(7):788–808, Oct. 2006.[229] J. Ortega-Arroyo and P. Kukura. Interferometric scattering microscopy(iSCAT): new frontiers in ultrafast and ultrasensitive optical microscopy.Physical Chemistry Chemical Physics, 14(45):15625–15636, 2012.[230] C. Pasquini. Near Infrared Spectroscopy: fundamentals, practical aspectsand analytical applications. Journal of the Brazilian Chemical Society,14:198 – 219, 04 2003.196[231] L. Pastia, B. Walczaka, D. Massarta, and P. Reschiglian. Optimization ofsignal denoising in discrete wavelet transform. Chemometrics andIntelligent Laboratory Systems, 48:21–34, 1999.[232] F. G. Pearson, R. H. Marchessault, and C. Y. Liang. Infrared spectra ofcrystalline polysaccharides. v. chitin. Journal of Polymer Science,43(141):101–116, Mar. 1960.[233] W. Petrich. Mid-infrared and raman spectroscopy for medical diagnostics.Applied Spectroscopy Reviews, 36(2-3):181–237, 2001.[234] R. Petry, M. Schmitt, and J. Popp. Raman spectroscopya prospective toolin the life sciences. ChemPhysChem, 4(1):14–30, 2003.[235] H. R. Petty. Fluorescence microscopy: Established and emerging methods,experimental strategies, and applications in immunology. MicroscopyResearch and Technique, 70(8):687–709, 2007.[236] J. L. Pichardo-Molina, C. Frausto-Reyes, O. Barbosa-Garcı´a,R. Huerta-Franco, J. L. Gonza´lez-Trujillo, C. A. Ramı´rez-Alvarado,G. Gutie´rrez-Jua´rez, and C. Medina-Gutie´rrez. Raman spectroscopy andmultivariate analysis of serum samples from breast cancer patients. Lasersin Medical Science, 22(4):229–236, 2007.[237] V. Piliarik, M. & Sandoghdar. Direct optical sensing of single unlabelledproteins and super-resolution imaging of their binding sites. NatureCommunications, 5:4495, 2014.[238] C. Pizarro, I. Esteban-Dez, J.-M. Gonzlez-Siz, and M. Forina. Use ofnear-infrared spectroscopy and feature selection techniques for predictingthe caffeine content and roasting color in roasted coffees. Journal ofAgricultural and Food Chemistry, 55:7477–7488, 2007.[239] J. S. Ploem. Reflection-contrast microscopy as a tool for investigation ofthe attachment of living cells to a glass surface . Mononuclear phagocytesin immunity, infection and pathology, pages 405–421, 1975.[240] J. S. Ploem. Laser scanning fluorescence microscopy. Applied Optics,26(16):3226–3231, Aug. 1987.[241] F. S. Poke and C. A. Raymond. Predicting extractives, lignin, and cellulosecontents using near infrared spectroscopy on solid wood in eucalyptusglobulus. Journal of Wood Chemistry and Technology, 26(2):187–199, July2006.197[242] F. S. Poke, J. K. Wright, and C. A. Raymond. Predicting extractives andlignin contents in eucalyptus globulus using near infrared reflectanceanalysis. Journal of Wood Chemistry and Technology, 24(1):55–67, Jan.2005.[243] E. O. Potma and X. S. Xie. Cars microscopy for biology and medicine.Opt. Photon. News, 15(11):40–45, Nov 2004.[244] E. O. Potma and X. S. Xie. CARS microscopy for biology and medicine.Optics and Photonics News, 15(11):40–45, Nov. 2004.[245] B. PR. Molecular Symmetry and Spectroscopy. Academic Press Inc.,,Newyork, 1979.[246] P. Pudney, T. Hancewicz, D. Cunningham, and M. Brown. Quantifying themicrostructures of soft solid materials by confocal raman spectroscopy.Vibrational Spectroscopy, 34(1):123–135, Jan. 2004.[247] G. J. Puppels, F. F. M. de Mul, C. Otto, J. Greve, M. Robert-Nicoud, D. J.Arndt-Jovin, and T. M. Jovin. Studying single living cells andchromosomes by confocal raman microspectroscopy. Nature,347:301–303, 1990.[248] V. I. Pustovoit, V. E. Pozhar, M. M. Mazur, V. N. Shorin, I. B. Kutuza, andA. V. Perchik. Double-aotf spectral imaging system. Proc. SPIE,5953:59530P–59530P–4, 2005.[249] C. V. Raman. A new radiation. Indian Journal of Physics, pages 387–389,1928.[250] C. V. Raman and K. S. Krishnan. A New Type of Secondary Radiation.nature, 121:501–502, Mar. 1928.[251] C. A. Raymond and L. R. Schimleck. Development of near infraredreflectance analysis calibrations for estimating genetic parameters forcellulose content in eucalyptus globulus. Canadian Journal of ForestResearch, 32(1):170–176, Jan. 2002.[252] C. A. Raymond, L. R. Schimleck, A. Muneri, and A. J. Michell.Nondestructive sampling of eucalyptus globulus and e. nitens for woodproperties. III. predicted pulp yield using near infrared reflectance analysis.Wood Science and Technology, 35(3):203–215, June 2001.198[253] E. G. Reynaud, U. Krzˇicˇ, K. Greger, and E. H. K. Stelzer. Lightsheet-based fluorescence microscopy: More dimensions, more photons,and less photodamage. HFSP Journal, 2(5):266–275, Oct. 2008.[254] H. T. R.H. Wilson. Mid-infrared spectroscopy for food analysis: recentnew applications and relevant developments in sample presentationmethods. TrAC Trends in Analytical Chemistry, 18(2):85 – 93, 1999.[255] E. F. J. Ring. The discovery of infrared radiation in 1800. The ImagingScience Journal, 48(1):1–8, 2000.[256] J. G. Ritter, R. Veith, A. Veenendaal, J. P. Siebrasse, and U. Kubitscheck.Light sheet microscopy for single molecule tracking in living tissue. PLoSONE, 5(7):11639, July 2010.[257] B. Robert. Resonance raman spectroscopy. Photosynthesis Research,101(2):147–155, 2009.[258] L. G. Rodriguez, S. J. Lockett, and G. R. Holtom. Coherent anti-stokesraman scattering microscopy: A biological review. Cytometry Part A,69A(8):779–791, 2006.[259] Y. Roggo, P. Chalus, L. Maurer, C. Lema-Martinez, A. Edmond, andN. Jent. A review of near infrared spectroscopy and chemometrics inpharmaceutical technologies. Journal of Pharmaceutical and BiomedicalAnalysis, 44(3):683 – 700, 2007.[260] G. D. Romdhane Karoui and C. Blecker. Mid-infrared spectroscopycoupled with chemometrics: A tool for the analysis of intact food systemsand the exploration of their molecular structure?quality relationships ? areview. Chemical Reviews, 10:6144–6168, 2010.[261] B. G. Saar, C. W. Freudiger, J. Reichman, C. M. Stanley, G. R. Holtom, andX. S. Xie. Video-rate molecular imaging in vivo with stimulated ramanscattering. Science, 330(6009):1368–1370, 2010.[262] P. A. Santi. Light sheet fluorescence microscopy: A review. Journal ofHistochemistry & Cytochemistry, 59(2):129–138, Feb. 2011.[263] B. A. SCALETTAR, J. R. SWEDLOW, J. W. SEDAT, and D. A. AGARD.Dispersion, aberration and deconvolution in multi-wavelength fluorescenceimages. Journal of Microscopy, 182(1):50–60, 1996.199[264] H. V. Scheller and P. Ulvskov. Hemicelluloses. Ann Rev Plant Biol,61:263–269, 2010.[265] K. Schenzel, H. Almlf, and U. Germga˚rd. Quantitative analysis of thetransformation process of cellulose i→ cellulose II using NIR FT ramanspectroscopy and chemometric methods. Cellulose, 16(3):407–415, Feb.2009.[266] K. Schenzel, S. Fischer, and E. Brendler. New method for determining thedegree of cellulose i crystallinity by means of FT raman spectroscopy.Cellulose, 12(3):223–231, June 2005.[267] L. Schermelleh, R. Heintzmann, and H. Leonhardt. A guide tosuper-resolution fluorescence microscopy. The Journal of Cell Biology,190(2):165–175, July 2010.[268] L. Schimleck and R. Evans. Estimation of air-dry density of incrementcores by near infrared spectroscopy. Appita Journal, 56:312–317, 2003.[269] L. R. Schimleck, P. D. Kube, and C. A. Raymond. Genetic improvement ofkraft pulp yield in eucalyptus nitens using cellulose content determined bynear infrared spectroscopy. Canadian Journal of Forest Research,34(11):2363–2370, Nov. 2004.[270] L. R. Schimleck, P. D. Kube, C. A. Raymond, A. J. Michell, and J. French.Estimation of whole-tree kraft pulp yield of eucalyptus nitens usingnear-infrared spectra collected from increment cores. Canadian Journal ofForest Research, 35(12):2797–2805, Dec. 2005.[271] T. Schmid, A. Messmer, B.-S. Yeo, W. Zhang, and R. Zenobi. Towardschemical analysis of nanostructures in biofilms II: tip-enhanced ramanspectroscopy of alginates. Analytical and Bioanalytical Chemistry,391(5):1907–1916, Apr. 2008.[272] B. Schrader. in Infrared and Raman Spectroscopy. VCH Weinheim, 2edition, 1995.[273] S. Schultz, D. R. Smith, J. J. Mock, and D. A. Schultz. Single-targetmolecule detection with nonbleaching multicolor optical immunolabels.Proceedings of the National Academy of Sciences, 97(3):996–1001, 2000.[274] T. Schultz and D. Burnds. Rapid secondary analysis oflignocelluloseo´comparison of near-infrared (nir) and fourier-transforminfrared (ftir). TAPPI Journal, 73:209–212, 1990.200[275] H. Schulz and M. Baranska. Identification and quantification of valuableplant substances by ir and raman spectroscopy. Vibrational Spectroscopy,43:13–25, 2007.[276] G. Seisenberger. Real-time single-molecule imaging of the infectionpathway of an adeno-associated virus. Science, 294:19291932, 2001.[277] J. Selinummi, P. Ruusuvuori, I. Podolsky, A. Ozinsky, E. Gold,O. Yli-Harja, A. Aderem, and I. Shmulevich. Bright field microscopy as analternative to whole cell fluorescence in automated analysis of macrophageimages. PLoS ONE, 4(10):1–9, Oct. 2009.[278] K. E. Shafer-Peltier, A. S. Haka, J. T. Motz, M. Fitzmaurice, R. R. Dasari,and M. S. Feld. Model-based biological raman spectral imaging. Journalof Cellular Biochemistry, 87(S39):125–137, 2002.[279] P. Shaw. Deconvolution in 3-d optical microscopy. The HistochemicalJournal, 26(9):687–694, 1994.[280] C. J. R. Sheppard, M. Gu, and M. Roy. Signal-to-noise ratio in confocalmicroscope systems. Journal of Microscopy, 168(3):209–218, 1992.[281] B. Sherman. Infrared spectroscopy by attenuated total reflection. Appl.Spectrosc., 18(1):7–9, Jan 1964.[282] R. Singh and F. Riess. The 1930 nobel prize for physics: A close decision?Notes and Records, 55(2):267–283, 2001.[283] V. Sinija and H. Mishra. Ft-nir spectroscopy for caffeine estimation ininstant green tea powder and granules. {LWT} - Food Science andTechnology, 42(5):998 – 1002, 2009.[284] E. Sjo¨holm, K. Gustafsson, E. Norman, T. Reitberger, and A. Colmsjo¨.Fibre strength in relation to molecular weight distribution of hardwoodkraft pulp: Degradation by gamma irradiation, oxygen/alkali or alkali.Nord. Pulp Pap. Res. J., 15:326–332, 2000.[285] E. Sjstrm. Wood chemistry: Fundamentals and applications. AcademicPress, 1993.[286] P. T. C. So, C. Y. Dong, B. R. Masters, and K. M. Berland. Two-photonexcitation fluorescence microscopy. Annual Review of BiomedicalEngineering, 2(1):399–429, Aug. 2000.201[287] D. Soliman, G. J. Tserevelakis, M. Omar, and V. Ntziachristos. Combinedlabel-free optical and optoacoustic imaging of model organisms atmesoscopy and microscopy resolutions, 2016.[288] J. S. Soul and A. J. du Plessis. Near-infrared spectroscopy. Seminars inPediatric Neurology, 6(2):101 – 110, 1999. New Technologies in PediatricNeurology.[289] Y. Sowa, B. C. Steel, and R. M. Berry. A simple backscattering microscopefor fast tracking of biological molecules. Rev. Sci. Instrum., 81:113704,2010.[290] H. L. Spiegelberg. The effect of Hemicelluloses on the mechanicalproperties of individual pulp fibers. PhD thesis, Lawrence University, 1966.[291] S. Spindler, J. Ehrig, K. Knig, T. Nowak, M. Piliarik, H. E. Stein, R. W.Taylor, E. Garanger, S. Lecommandoux, I. D. Alves, and V. Sandoghdar.Visualization of lipids and proteins at high spatial and temporal resolutionvia interferometric scattering (iscat) microscopy. Journal of Physics D:Applied Physics, 49(27):274002, 2016.[292] T. G. Spiro. Resonance raman spectroscopy. new structure probe forbiological chromophores. Accounts of Chemical Research, 7(10):339–344,Oct. 1974.[293] B. Stallinga, S. & Rieger. Accuracy of the gaussian point spread functionmodel in 2d localization microscopy. Opt. Express, 18:2446124476, 2010.[294] S. Stallinga and B. Rieger. Accuracy of the gaussian point spread functionmodel in 2d localization microscopy. Opt. Express, 18(24):24461–24476,Nov 2010.[295] D. J. Stephens and V. J. Allan. Light microscopy techniques for live cellimaging. Science, 300(5616):82–86, 2003.[296] P. L. Stiles, J. A. Dieringer, N. C. Shah, and R. P. V. Duyne.Surface-enhanced raman spectroscopy. Annual Review of AnalyticalChemistry, 1(1):601–626, July 2008.[297] D. P. Strommen and K. Nakamoto. Resonance raman spectroscopy.Journal of Chemical Education, 54:474, 1977.[298] W. S. Struve. Fundamentals of Molecular Spectroscopy. McGraw-HillBook Company (UK) Limited, 1989.202[299] B. Stuart. Infrared Spectroscopy. John Wiley & Sons, Inc., 2000.[300] G. E. Stutzmann. Dynamic multiphoton imaging: A live view from cells tosystems. Physiology, 20(1):15–21, Feb. 2005.[301] J. Sun. Statistical analysis of nir data: Data pretreatment. J Chemometr.,11:525–532, 1997.[302] Z. Sun, A. Ibrahim, P. B. Oldham, T. P. Schultz, and T. E. Conners. Rapidlignin measurement in hardwood pulp samples by near-infrared Fouriertransform raman spectroscopy. J. Agric. Food Chem., 45(8):3088–3091,Aug. 1997.[303] D. Sundararajan. Discrete Wavelet Transform: A Signal ProcessingApproach. 2015.[304] J. Sutter and J. Kalivas. Comparison of forward selection, backwardelimination, and generalized simulated annealing for variable selection.Microchemical Journal, 47:60–66, 1993.[305] E. Suzuki. High-resolution scanning electron microscopy ofimmunogold-labelled cells by the use of thin plasma coating of osmium.Journal of Microscopy, 208(3):153–157, Dec. 2002.[306] K. Svoboda and S. M. Block. Biological applications of optical forces.Annual Review of Biophysics and Biomolecular Structure, 23(1):247–285,June 1994. PMID: 7919782.[307] N. Tavassoli, Z. Chen, A. Bain, L. Melo, D. Chen, and E. R. Grant.Template-oriented genetic algorithm feature selection of analyte waveletsin the raman spectrum of a complex mixture. Anal. Chem.,85:10591–10599, 2014.[308] N. Tavassoli, W. Tsai, P. Bicho, and E. R. Grant. Multivariate classificationof pulp NIR spectra for end-product properties using discrete wavelettransform with orthogonal signal correction. Anal. Methods,6(22):8906–8914, July 2014.[309] D. Tavast, Z. A. Mansoor, and E. Brnnvall. Xylan from agro waste as astrength enhancing chemical in kraft pulping of softwood. Ind. Eng. Chem.Res., 53:9738–9742, 2014.[310] L. G. Thygesen and S.-O. Lundqvist. Nir measurement of moisture contentin wood under unstable temperature conditions: part 2. handling203temperature fluctuations. Journal of Near Infrared Spectroscopy,8:191–199, 2000.[311] Z. Q. Tian. Surface-enhanced raman spectroscopy: advancements andapplications. Journal of Raman Spectroscopy, 36(6-7):466–470, 2005.[312] P. Tinnefeld and M. Sauer. Branching out of single-molecule fluorescencespectroscopy: Challenges for chemistry and influence on biology.ChemInform, 36(31):2642–2671, Aug. 2005.[313] H. G. Tompkins. The physical basis for analysis of the depth of absorbingspecies using internal reflection spectroscopy. Appl. Spectrosc.,28(4):335–341, Jul 1974.[314] H. Tong, M. Li, H. Zhang, and C. Zhang. Blur detection for digital imagesusing wavelet transform. In 2004 IEEE International Conference onMultimedia and Expo (ICME) (IEEE Cat. No.04TH8763). Institute ofElectrical and Electronics Engineers (IEEE), 2004.[315] S. Tsuchikawa. A review of recent near infrared research for wood andpaper. Applied spectroscopy reviews, 42:43–71, 2007.[316] S. Tsuchikawa, Y. Hirashima, Y. Sasaki, and K. a. Ando. Near-infraredspectroscopic study of the physical and mechanical properties of wood withmeso- and micro-scale anatomical observation. Applied Spectroscopy,59:86–93, 2005.[317] S. Tsuchikawa and S. Tsutsumi. Adsorptive and capillary condensed waterin biological material. JOURNAL OF MATERIALS SCIENCE LETTERS,17:661–663, 1998.[318] S. Tsuchikawa and S. Tsutsumi. Application of near infraredspectrophotometry to wood. Applied Spectroscopy, 53:1033–1039, 1999.[319] T. L. U. and Edlund. Prediction of lignin content and pulp yield from blackliquor composition using near-infrared spectroscopy and partial leastsquares regression. Nordic Pulp & Paper Research Journal, 13:76–80,1998.[320] S. A. R. Umesh P. Agarwal, Richard R. Reiner. Estimation of cellulosecrystallinity of lignocelluloses using near-IR FT-raman spectroscopy andcomparison of the raman and segal-WAXS methods. J. Agric. Food Chem.,61(1):103–113, Jan. 2013.204[321] S. A. R. Umesh P. Agarwal, Richard S. Reiner. Cellulose i crystallinitydetermination using FT–raman spectroscopy: univariate and multivariatemethods. Cellulose, 17(4):721–733, May 2010.[322] Y. Urano, M. Sakabe, N. Kosaka, M. Ogawa, M. Mitsunaga, D. Asanuma,M. Kamiya, M. R. Young, T. Nagano, P. L. Choyke, and H. Kobayashi.Rapid cancer detection by topically spraying a-glutamyltranspeptidase–activated fluorescent probe. Science TranslationalMedicine, 3(110):110–119, Nov. 2011.[323] M. Urbano-Cuadrado, M. L. de Castro, P. Prez-Juan, J. Garca-Olmo, andM. Gmez-Nieto. Near infrared reflectance spectroscopy and multivariateanalysis in enology: Determination or screening of fifteen parameters indifferent types of wines. Analytica Chimica Acta, 527(1):81 – 88, 2004.[324] M. Valcarcel, M. D. L. de Castro, and M. T. Tena. Analytical viewpoint.preliminary operations: a pending goal of today’s analytical chemistry.Anal. Proc., 30:276–279, 1993.[325] P. J. Verveer, J. Swoger, F. Pampaloni, K. Greger, M. Marcello, andE. H. K. Stelzer. High-resolution three-dimensional imaging of largespecimens with light sheet–based microscopy. Nature Methods, 4:311–313,Mar. 2007.[326] J. Vester, C. Felby, O. F. Nielsen, and S. Barsberg. Fourier transform ramandifference spectroscopy for detection of lignin oxidation products inthermomechanical pulp. Applied Spectroscopy, 58(4):404–409, Apr. 2004.[327] B. Via, T. Shupe, L. Groom, M. Stine, and C. So. Multivariate modelling ofdensity and strength and stiffness from near infrared spectra for mature,juvenile and pith wood of longleaf pine. Journal of Near InfraredSpectroscopy, 11:365–378, 2003.[328] B. Walczak. Wavelets in Chemistry. Elsevier, 2000.[329] B. Walczak and D. Massart. Wavelets o´ something for analyticalchemistry? TrAC Trends in Analytical Chemistry, 16:451–463, 1997.[330] C.-S. Wang. Theory of stimulated raman scattering. Phys. Rev.,182:482–494, Jun 1969.[331] G. Wang, A. S. Stender, W. Sun, and N. Fang. Optical imaging ofnon-fluorescent nanoparticle probes in live cells. The Analyst,135(2):215–221, 2010.205[332] Z. Wang, J. Feng, L. Li, W. Ni, and Z. Li. A non-linearized pls modelbased on multivariate dominant factor for laser-induced breakdownspectroscopy measurements. J. Anal. At. Spectrom., 26:2175–2182, 2011.[333] L. Wei, Y. Yu, Y. Shen, M. C. Wang, and W. Min. Vibrational imaging ofnewly synthesized proteins in live cells by stimulated raman scatteringmicroscopy. Proceedings of the National Academy of Sciences,110(28):11226–11231, 2013.[334] N. Wei, E. Flaschel, K. Friehs, and T. Nattkemper. A machine visionsystem for automated non-invasive assessment of cell viability via darkfield microscopy, wavelet feature selection and classification. BMCBioinformatics, 9(1):449, 2008.[335] N. Wei, J. You, K. Friehs, E. Flaschel, and T. W. Nattkemper. In situ darkfield microscopy for on-line monitoring of yeast cultures. BiotechnologyLetters, 29(3):373–378, Dec. 2006.[336] S. Weiss. Fluorescence spectroscopy of single biomolecules. Science,283(5408):1676–1683, Mar. 1999.[337] Z. Q. Wen, L. D. Barron, , and L. Hecht. Vibrational raman optical activityof monosaccharides. J. Am. Chem. SOC, 115:285–292, 1993.[338] L. G. Weyer. Near-infrared spectroscopy of organic substances. AppliedSpectroscopy Reviews, 21(1-2):1–43, 1985.[339] M. Whittle, V. J. Gillet, P. Willett, and J. Loesel. Analysis of data fusionmethods in virtual screening: similarity and group fusion. Journal ofChemical Information and Modeling, 46(6):2206–2219, 2006. PMID:17125165.[340] J. H. Wiley and R. H. Atalla. Band assignments in the raman spectra ofcelluloses. Carbohydrate Research, 160:113–129, Feb. 1987.[341] P. Willett. Combination of similarity rankings using data fusion. Journal ofChemical Information and Modeling, 53(1):1–10, 2013. PMID: 23297768.[342] E. Wilson, J. Decius, and P. Cross. Molecular Vibrations. McGraw-Hill,New York, 1955.[343] E. Windeisen and G. Wegener. Behaviour of lignin during thermaltreatments of wood. Industrial Crops and Products, 27(2):157–162, Mar.2008.206[344] N. Wistara and R. A. Young. Properties and treatments of pulps fromrecycled paper. part i. physical and chemical properties of pulps. Cellulose,6(4):291–324, 1999.[345] A. Wojciak, H. Kasprzyk, I. Khmelinskii, A. Krawczyk, A. Oliveira,L. Ferreira, A. Weselucha-Birczynska, and M. Sikorski. Directcharacterization of hydrogen peroxide bleached thermomechanical pulpusing spectroscopic methods. J Phys Chem A, 111:10530–10536, 2007.[346] D. L. Wokosin, V. F. Centonze, J. G. White, S. N. Hird, S. Sepsenwol,G. P. A. Malcolm, G. T. Maker, and A. I. Ferguson. Multiple-photonexcitation imaging with an all-solid-state laser. In D. L. Farkas, R. C. Leif,A. V. Priezzhev, T. Asakura, and B. J. Tromberg, editors, OpticalDiagnostics of Living Cells and Biofluids, volume 2678. SPIE-Intl SocOptical Eng, May 1996.[347] S. Wold, H. Antti, F. Lindgren, and J. Ohman. Orthogonal signal correctionof near-infrared spectra. Chemometrics and Intelligent LaboratorySystems, 44:175–185, 1998.[348] S. Wold, M. Sjstrm, and L. Eriksson. Pls-regression: a basic tool ofchemometrics. Chemometrics and Intelligent Laboratory Systems,58:109–130, 2001.[349] J. J. Workman. Infrared and Raman spectroscopy in paper and pulpanalysis. Applied Spectroscopy Reviews, 36(2-3):139–168, June 2001.[350] J. Wright, M. Birkett, and M. Gambino. Prediction of pulp yield andcellulose content from wood samples using near-infrared reflectancespectroscopy. TAPPI Journal, 73:164–166, 1990.[351] Z. Wu, M. Du, C. Sui, B. Xu, Y. Peng, X. Shi, and Y. Qiao. Developmentand validation of a portable aotf-nir measurement method for thedetermination of baicalin in yinhuang oral solution. In 2012 InternationalConference on Biomedical Engineering and Biotechnology, pages1322–1326, May 2012.[352] Q.-S. Xua and Y.-Z. Liang. Monte carlo cross validation. Chemometricsand Intelligent Laboratory Systems, 56:1–11, 2001.[353] H. Yang, L. Xu, K. Chen, X. Huang, Q. He, and G. Jin. Improvement ofminiature grating spectrometers, 2007.207[354] B.-S. Yeo, J. Stadler, T. Schmid, R. Zenobi, and W. Zhang. Tip-enhancedraman spectroscopy – its status, challenges and future directions. ChemicalPhysics Letters, 472(1-3):1–13, Apr. 2009.[355] P. R. Yildiz, A. & Selvin. Fluorescence imaging with one nanometeraccuracy: Application to molecular motors. Acc. Chem. Res., 38:574582,2005.[356] G. Yu-mei and Z. Wei. Recent progress in nir spectroscopy technology andits application to the field of forestry. Spectroscopy and Spectral Analysis,28:1544–1548, 2008.[357] G. W. Zack, W. E. Rogers, and S. A. Latt. Automatic measurement of sisterchromatid exchange frequency. The Journal of Histochemistry andCytochemistry., 25:741–753, 1977.[358] S. Zhang, H. Gao, and G. Bao. Physical principles of nanoparticle cellularendocytosis. ACS Nano, 9(9):8655–8671, Sept. 2015.[359] R. G. Zhbankov. Vibrational spectra and structure of mono- andpolysaccharides. Journal of Molecular Structure, 275:65–84, 1992.[360] G. Zhou, G. Taylor, and A. Polle. FTIR-ATR-based prediction andmodelling of lignin and energy contents reveals independent intra-specificvariation of these traits in bioenergy poplars. Plant Methods, 7(1), 2011.[361] X. Zhuang. Nano-imaging with STORM. Nature Photonics, 3(7):365–367,July 2009.[362] A. Zieba, C. Wahlby, F. Hjelm, L. Jordan, J. Berg, U. Landegren, andK. Pardali. Bright-field microscopy visualization of proteins and proteincomplexes by in situ proximity ligation with peroxidase detection. ClinicalChemistry, 56(1):99–110, Nov. 2009.[363] W. R. Zipfel, R. M. Williams, R. Christie, A. Y. Nikitin, B. T. Hyman, andW. W. Webb. Live tissue intrinsic emission microscopy usingmultiphoton-excited native fluorescence and second harmonic generation.Proceedings of the National Academy of Sciences, 100(12):7075–7080,May 2003.[364] T. Zu¨chner, A. V. Failla, M. Steiner, and A. J. Meixner. Probing dielectricinterfaces on the nanoscale with elastic scattering patterns of single goldnanorods. Opt. Express, 16(19):14635–14644, Sep 2008.208[365] A. Zumbusch, G. R. Holtom, and X. S. Xie. Three-dimensional vibrationalimaging by coherent anti-stokes raman scattering. Phys. Rev. Lett.,82:4142–4145, May 1999.209Appendix A:Physical Properties of PulpFigure A.1: The correlation between different mechanical properties of pulpsheets made of a slightlybeaten pulp with freeness 600210Table A.1: The experimental range, average and the standard deviation ofmeasurment of different physical- mechanical properties of pulp sam-ples that served as calibration and validation standards in chapter 3-5.character Range average median Standard deviationFreeness 649.50 - 700.50 656.01 700.50 12.11Tensile initial 2.50-4.87 3.77 3.76 0.54Tensile 600 4.30-7.67 6.19 6.21 0.68Tensile 500 6.50-9.36 8.27 8.35 0.70Tensile 450 7.91 - 10.63 9.38 9.43 0.65Tensile 300 9.20 - 11.56 10.41 10.44 0.51Tear initial 19.86 - 30.34 26.18 26.64 1.93Tear 600 17.34 - 24.29 20.51 20.55 1.39Tear 500 12.15 - 20.13 15.62 15.43 1.63Tear 450 5.76 - 8.37 13.28 13.20 1.65Tear 300 8.25 - 15.22 11.40 11.47 1.34Burst inital 1.36 - 3.37 2.33 2.35 0.43Burst 600 3.20 - 5.95 4.42 4.40 0.53Burst 500 4.90 - 7.63 6.25 6.27 0.56Burst 450 5.76 - 8.37 7.22 7.26 0.52Burst 300 6.55 - 9.18 8.10 8.14 0.42SRE 600 23.91 - 82.73 50.69 51.91 14.44SRE 500 60.81 - 162.50 111.57 115.45 23.52SRE 450 92.35 - 217.97 157.59 161.25 29.29SRE 300 152.90 - 309.08 236.65 239.97 36.08Density 0.47 - 0.57 0.53 0.53 0.02Density 600 0.54 - 0.61 0.58 0.58 0.01Density 500 0.59 - 0.65 0.62 0.62 0.01Density 450 0.62 - 0.67 0.65 0.65 0.01Density 300 0.64 - 0.70 0.68 0.68 0.01Wet zero span 12.40 - 15.82 14.59 14.69 0.60Dry zero span 14.42 - 16.78 15.65 15.69 0.53Absorption 0.1669 - 0.227 0.19 0.19 0.01Scattering 32.6 - 39.8 35.50 35.57 1.51Scattering 600 29.10 - 35.37 31.47 31.45 1.29Scattering 500 25.52 - 31.47 27.98 27.97 1.11Scattering 450 24.10 - 29.21 26.16 26.16 1.00Scattering 300 22.35 - 26.80 24.38 24.31 0.96211Appendix B:A Comparative Study of Different Machine Learning Techniques In Predict-ing Properties of Paper ProductsIn this section, we build machine learning models to predict the physical prop-erties of pulp products using Raman spectra of pulp sample. As we discussed inchapter 4, the data is high dimensional, and the number of samples is limited toless than hundred. So, we used a feature selection and reduction techniques beforeapplying prediction model.In supervised machine learning, a setup data labeled with target variables areused to train a regression model. To evaluate and improve the model, the data setis divided into training and validation sets. The learner fit a regression model usingthe training data established in a way that performs well in predicting the targetvariables in the validation set.Next, we utilized feature extraction/selection techniques to find most importantfeatures representing the input pattern to decrease the data dimensions. Feature ex-traction/selection techniques are considered as a preprocessing step of data miningresult in a less computational effort and in improving the accuracy of the model.In table B.1 we report root mean square error obtained from all the regressorswhen they used the selected and extracted features together. Comparing this tablewith the TOGA results presented in the chapter 5 shows that TOGA improves pre-diction models for all mechanical properties of pulpsheets more efficiently. Thisimprovement proves that this method extracts reconstructed spectra that offer im-portant and sufficient information for classification.212Table B.1: Root mean square error all the regressors when they used the se-lected features and extracted features together.Method Average KNN Ridge Lasso LassoLars Bayesian RidgeFreeness 61.1 12.7 8.1 8.7 8.9 9.5Initial Tensile 4.8 0.81 0.63 0.68 0.65 0.62Tensile600 3.9 0.69 0.57 0.57 0.57 0.59Tensile 550 4.3 0.72 0.57 0.58 0.62 0.57Tensile 500 4.1 0.77 0.53 0.53 0.56 0.56Tensile 450 4.28 0.74 0.55 0.55 0.55 0.54Tensile 300 4.2 0.72 0.53 0.56 0.57 0.52Initial Tear 3.7 2.7 1.9 2.1 2.1 2.2Tear600 3.2 1.5 0.98 0.97 0.99 1.1Tear550 3.7 1.7 1.2 1.1 1.1 1.3Tear500 4.8 2.5 1.7 2.1 2.3 1.5Tear450 4.2 2.7 1.5 1.5 1.6 1.7Tear300 4.8 2.1 1.2 1.2 1.4 1.5Initial Burst 4.3 0.63 0.40 0.32 0.32 0.47Burst600 4.0 0.64 0.46 0.47 0.52 0.48Burst550 4.1 0.67 0.44 0.46 0.43 0.50Burst500 4.5 0.67 0.53 0.53 0.53 0.49Burst450 4.2 0.65 0.46 0.47 0.50 0.46Burst400 4.1 0.55 0.40 0.40 0.36 0.39SRE600 11.4 15.0 10.9 11.1 11.8 10.1SRE550 19.4 21.7 16.6 18.4 18.4 13.6SRE500 23.8 25.8 20.2 19.6 19.6 17.2SRE450 33.1 32.9 27.7 27.8 27.9 22.7SRE300 47.8 45.5 35.2 34.7 34.7 36.3Initial Density 3.9 0.029 0.017 0.017 0.020 0.021Density600 3.8 0.021 0.014 0.014 0.014 0.014Density550 3.9 0.017 0.013 0.013 0.014 0.014Density500 3.9 0.017 0.012 0.012 0.012 0.011Density450 4.1 0.015 0.011 0.009 0.010 0.010Density300 4.1 0.015 0.013 0.013 0.013 0.010Wet zero span 3.9 0.74 0.66 0.67 0.69 0.65Dry zero span 3.7 0.62 0.46 0.46 0.45 0.46Absorption 3.9 0.019 0.0088 0.0082 0.0085 0.0086Scattering 2.5 2.28 1.5 1.4 1.4 1.8213Appendix C:iSCAT Application 1, iSCAT Images of Single Fibre and Pulp SamplesFigure C.1: A deconvolved iSCAT image of a single fibre in a pulp sample214Figure C.2: DWT deconvolved iSCAT images of a single fiber in different stacks containing informationof 24 micron in Z direction, 20 micron in X and 60 micron in Y direction.215Figure C.3: A reconstructed volume using the deconvolved stacks in Figurefig:fibre.216Figure C.4: DWT deconvolved iSCAT images of a single fiber in different stacks containing informationof 20 micron in Z direction, 20 micron in X and 60 micron in Y direction.217Figure C.5: A reconstructed volume using the deconvolved stacks in Figurefig:fibre2.218Figure C.6: DWT deconvolved iSCAT images of a pulp sample representingthe network of fibres in different stacks containing information of 0micron in Z direction, 60 micron in X and 60 micron in Y direction.219Appendix D:iSCAT Application 2, Fluid Trapped in OlivineThis work has been done in collaboration with Department of Earth, Oceanand Atmospheric Sciences at UBC. The sample is Olivine (Mg2SiO4) contains lit-tle cavities in it. The aim of our collaboration was to detect these cavities andinvestigate if they are filled with fluid or not. The composition of this fluid wasanother scope of this study.Figure D.1: 3D iSCAT image of an olivin sample220Figure D.1 presents an iSCAT image of one olivine sample. The red part isolivine, and blue-green color represents cavities in the sample. The size of thesecavities is around 10*10 micron in the X-Y plane. The 3D iSCAT image showsthat these holes are connected underneath of the surface.Figure D.2: Raman spectra of an olivine sample contains fluid. Green spec-trum shows the Raman spectrum of olivine while red and blue spectrarepresent area of the sample with a hole. We were interested to find thecomposition of the fluid which fills up the hole. The results show thatit is possible that the hole is empty or the liquid has a low Raman crosssection which hasn’t been shown up in Raman spectrum.221Figure D.3: left figure shows iSCAT image of an olivine sample with olivinein red and holes in blue and yellowish colors. right figure representsRaman map of around the same area of the olivine sample which weused to take iSCAT image. PC1 was used to build the present Ramanmap.222Figure D.4: Left shows the iSCAT image of surface of an olivine sample.Right figure represent the same area of olivine sample in a stack with 5micron deeper. It shows that the holes in surface get smaller in deeperregions223Appendix E:iSCAT Aplication 3, Study of Spinal CordFigure E.1: This image shows the surface of a sample of mouse’s spinal cordrepresenting a neuron.224Appendix F:iSCAT Application 4, Biomineralization: iSCAT/Raman Analysis of Mineral-ized ElastinMineralisation of extracellular matrix is essential in the development and func-tion of the skeletal system. Similar mineralization is observed in soft tissues underpathological conditions and is particularly problematic when seen in the vascularsystem. Our studies are primarily looking into the role of elastin (and collagen)degradation in the progression of vascular calcification. They could indicate thatdigestion of elastin with cathepsin K, a vital protease primarily secreted from os-teoclasts and macrophages, increases the accumulation of mineralized plaques inthe elastin, presumably due to an increase in available nucleation sites.Mineralisation of elastin fragments occurs by different mechanisms associatedwith available nucleation sites, and digestion with catK may increase the avail-ability of certain nucleation leading to differences in quantity and mechanism ofmineralization.Live interferometric scattering provides ultrafast and high resolution opticalimaging to determine the composition of a substance at a molecular level.Figure F.1: iSCAT images of different parts of an elastin sample.225Figure F.2: iSCAT images of an elastin sample with the calcificied crystals.226Figure F.3: iSCAT images of an elastin sample with the calcificied crys-tals, the calcificied crystals are shown up in brighter color comparedto elastin. 227Figure F.4: Top figure shows the iSCAT image of an elastin sample with cal-cificied crystals, the bottom image serves Raman map of the same sam-ple around the same region which iSCAT image has been taken. PC1was used to make Raman map.228Appendix G:iSCAT Application 5, Visualization of PDA-Primed Silicone SurfacesBased on recent discoveries related to the methods used by mussels and othermarine organisms to stick to surfaces (e.g., rocks) in the marine environment, therehas been interest in a bio-inspired approach to modifying material surfaces. Poly-dopamine (PDA) appears to be involved in such binding mechanisms and, as foundby others, PDA can not only bind strongly to both inorganic and organic surfacesbut also provides a reactive surface for complexation of other molecules (e.g., moi-eties with thiol or amine groups, metals, certain drugs). We therefore had an in-terest in using PDA as a primer to attach drugs which might then be released in acontrolled manner either initially in vitro or eventually in vivo.This study was in collaboration with Dr. David Plackett, from Pharmaceuti-cal Sciences, UBC. The work in pharmaceutical department was supported by theNSERC CREATE SusSyn network in the form of a summer student dedicated tothe project in summer 2015 and focused on PDA priming of silicone films as amodel material. We were then interested in studying the binding and release of theaminoglycoside antibiotic gentamicin. A series of studies were run in which therelease of gentamicin from PDA-primed silicone was characterised.iSCAT was used to visualize of PDA-primed silicone surfaces before and afterloading with gentamicin. A series of images and videos were obtained and theseillustrated both the thickness of the PDA coating and the surface morphology.229Figure G.1: This image shows PDA-primed silicone surfaces. The thicknessof PDA coating and the surface morphology are easily can be detected230Figure G.2: This image shows PDA-primed silicone surfaces loaded withgentamicin. Comparing between this figure and above figure showsthe different surface morphology before and after loading drug.231Appendix H:iSCAT Application 6, Visualization of Biological CellsFigure H.1: A DWT deconvolved iSCAT image of heLa cell using the stan-dard deviation of 25 Z stacks.232Figure H.2: DWT deconvolved iSCAT images of a heLa cell sample in fifteen different depth.233Appendix I:MATLAB CodesThe developed MATLAB codes used in this thesis have been uploaded to my”GitHub” as follows:Data Mining in Analytical ChemistryThe supplementary documents such as iSCAT videos could be found in:Online Study of Various Complex Sample.234


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items